SDK & Tools

Python
TypeScript
CLI
Plugins

v2.2.1

Bug Fixes:

Core: add() (Memory and AsyncMemory) no longer reports records the vector store rejected as successful ADD events. Only records that were actually inserted are written to history, entity-linked, and returned. If none of the extracted memories could be inserted, add() now raises VectorStoreError instead of returning memories that were never stored (#7066)
Core: Restore the context-manager protocol on Memory (with Memory() as m:) and AsyncMemory (async with AsyncMemory() as m:), which closes the instance on exit (#7354)
LLMs: The AWS Bedrock Anthropic path now returns the first Converse content block that carries text instead of always reading content[0]. Claude reasoning models can emit a reasoningContent block before the text block, which previously made the call fail with a KeyError (#6369)
Vector Stores: Turbopuffer filters now apply every operator (eq, ne, gt, gte, lt, lte, in, nin). Only gte and lte were read before, so any other operator was silently dropped and the query returned unfiltered results. An unsupported operator now raises ValueError (#6564)
Vector Stores: Turbopuffer search() scores now respect distance_metric. With euclidean_squared, the unbounded squared distance maps to 1 / (1 + distance) instead of 1 - distance, which went negative and inverted ranking for any distance above 1 (#6559)
Vector Stores: S3 Vectors search() scores are now metric-aware. With euclidean, the distance maps to 1 / (1 + distance) instead of max(0, 1 - distance), which collapsed most scores to 0 (#6547)

v2.2.0

New Features:

Client: Add User Profiles to MemoryClient and AsyncMemoryClient: get_profile(), generate_profile(), get_profile_settings(), update_profile_settings(), sample_profiles(), and get_profile_job(). A profile is a structured, always-current JSON summary of one user, shaped by a JSON Schema you configure per project and filled by an LLM from that user’s memories. Generation is asynchronous. Every job POST carries an Idempotency-Key; to retry a lost request without starting a second job, pass the same idempotency_key on each attempt (#7340)

Bug Fixes:

Vector Stores: Guard against None timestamps in the Valkey vector store’s insert() and update() paths. created_at and updated_at fields that were present in the payload but set to None previously passed the "created_at" not in payload / "updated_at" in payload checks and raised TypeError when datetime.fromisoformat() received None. Both paths now use .get() with a truthiness check so None values fall through to the default, matching the Redis provider’s behavior (#6993)

v2.1.0

Improvements:

Client: Requests now carry three surface-identity headers so the platform can tell which product made a call. X-Mem0-Source names the surface and X-Application the host app it runs inside, both set-once so a wrapper that already declared its identity keeps it. X-Mem0-Client is append-only and carries name/version per layer, outermost first, so a plugin calling this SDK reports the whole chain rather than only the last speaker. MEM0_SOURCE, MEM0_APPLICATION and MEM0_CLIENT_STACK set them from the environment for wrappers that cannot pass options (#7326)
Client: The client stack is bounded by dropping whole entries rather than slicing characters, and this SDK’s own entry is the reserved one. Truncating the joined string could sever an identifier mid-name and the platform parsed the fragment as a real client (#7326)

v2.0.20

Improvements:

OSS notices: Notice configuration now comes from a static, cacheable repository file with a bundled disabled fallback and deterministic rollout assignment, instead of calling PostHog’s feature-flag evaluation API. This keeps notices fail-safe when the remote config is unavailable and removes the PostHog feature-flag request from notice evaluation (#7185)
Vector Stores: RedisDBConfig now uses Pydantic’s native extra="forbid" handling for unknown fields instead of a custom model validator, preserving strict validation while returning standard Pydantic errors (#7089)

v2.0.19

Bug Fixes:

Embeddings: HuggingFaceEmbedding now falls back to the HUGGINGFACE_API_KEY env var, then a placeholder key, when huggingface_base_url is set and no api_key is configured. The OpenAI-compatible client used to talk to TEI endpoints raises at construction when no key resolves at all, so a TEI deployment that doesn’t require a real key previously failed to initialize (#6947)
Core: remove_code_blocks() now accepts list-shaped content (a sequence of {"text": ...} blocks, as some agent frameworks pass) by joining each block’s text before stripping code fences, instead of raising AttributeError from calling .strip() on a list (#6947)
Core: create_procedural_memory() (Memory and AsyncMemory) now raises a clear ValueError when the LLM returns no content for the summary, instead of continuing with empty content that surfaced as a confusing error further down the call (#6947)
Proxy: mem0.proxy no longer auto-installs litellm via a pip install subprocess when the import fails; it now raises ImportError with instructions to install it yourself. The auto-install could hang or fail silently in restricted environments and ran an unreviewed install on the caller’s behalf (#6947)
Client: get_all() (sync and async) now sends page and page_size as independent query params instead of requiring both to be set before either was sent. Passing only page_size without page previously had it silently dropped, so results came back at the server’s default page size (#6900)
LLMs: Add provider_override to AWSBedrockConfig, an explicit provider name (for example "anthropic") for when model is an application inference profile ARN whose opaque ID has no provider substring for extract_provider() to detect. Without it, those ARNs raised ValueError: Unable to determine provider (#6899)
LLMs: VllmConfig now falls back to the VLLM_BASE_URL env var when vllm_base_url isn’t passed explicitly. The default was filled in before the env var was ever checked, so VLLM_BASE_URL was silently ignored (#6897)

v2.0.18

Bug Fixes:

Core: Percent-escape %, &, and = in user_id, agent_id, and run_id when building the session scope key for the recent-conversation buffer, so the key stays unambiguous for ids containing those characters. Ordinary ids keep their existing key; an id already containing %, &, or = maps to a new key, so its buffer starts empty once and refills on the next add(). Stored memories are unaffected (#6892)
Vector Stores: Raise ValueError when a PGVector in/nin filter value is not a list. A string value was previously iterated character by character into the generated = ANY(...) array (so {"user_id": {"in": "alice"}} matched a, l, i, c, e), and a non-iterable value raised a bare TypeError from deep inside filter building (#6879)
Vector Stores: Reject index_accuracy=0 in the Oracle AI Vector Search config. The range check sat behind a truthiness test, so 0 skipped validation entirely and was passed through to WITH TARGET ACCURACY 0 instead of raising (#6848)
Vector Stores: Close the Oracle connection or pool that Mem0 opened when initialization fails. A client version check, a database version check, or a create_col() error previously propagated with the connection still open, leaking it for the life of the process. A caller-supplied client is left untouched (#6839)

v2.0.17

New Features:

Client: Add agent_custom_instructions to project.update()/update_project() (sync and async) and to the ProjectUpdateOptions and AddMemoryOptions typed models. It sets a second extraction instruction set that applies only to agent-scoped memories: an add passing agent_id without user_id uses it, one passing both splits by attribution, and while it is unset custom_instructions continues to apply to every memory (#6809)

v2.0.16

New Features:

Client: Add reference_date, latest_only, and keyword_search to SearchMemoryOptions, and latest_only to GetAllMemoryOptions, keeping the Python client’s typed options in sync with the Platform API and the CLIs (#6696)

Bug Fixes:

Core: Stop add() metadata from setting a memory’s identity scope. _build_filters_and_metadata() now strips user_id, agent_id, run_id, and actor_id from caller-supplied metadata before building the creation template, so metadata can no longer place a memory into a scope that was never passed through the entity params (#6656)
Vector Stores: Validate Upstash filter keys and values in search(), keyword_search(), and list(). Filter keys must match a safe identifier pattern, values must be str/int/float/bool, and string values containing a double quote or backslash are now rejected instead of being interpolated unescaped into the generated query string (#5981)
Embeddings: FastEmbedEmbedding.embed() now converts its result with .tolist() before returning, so callers get a plain List[float] instead of a numpy array (#6770)
Embeddings: HuggingFaceEmbedding now passes api_key to the OpenAI-compatible client when huggingface_base_url is set. The configured key was previously dropped, so the client fell back to OPENAI_API_KEY from the environment or raised OpenAIError at construction when that was unset (#6770)
Embeddings: Replace Ollama’s interactive pip install prompt on import with a plain ImportError. Importing mem0.embeddings.ollama without the ollama package previously blocked on stdin and then called sys.exit(1), killing the host process instead of raising (#6770)
Core: remove_code_blocks() now returns an empty string for None input instead of raising AttributeError (#6770)
Core: parse_vision_messages() now chains the original exception (raise ... from e) when an image download fails, so the root cause is preserved in the traceback (#6770)
Core: process_telemetry_filters(None) now returns ([], {}), matching the two-value tuple every caller unpacks, instead of {} (#6770)
Core: LlmFactory.create() no longer mutates the caller’s config dict in place via .update(kwargs); the merge now builds a new dict (#6770)
Rerankers: The Cohere, HuggingFace, SentenceTransformer, and Zero Entropy rerankers’ failure-fallback path no longer mutates the caller’s document dicts in place when stamping rerank_score; it now falls back on copies (#6770)
Vector Stores: Remove logging.basicConfig() calls from the MongoDB and Vertex AI Vector Search providers, so selecting either provider no longer reconfigures the host application’s root logger as a side effect (#6770)

v2.0.15

Bug Fixes:

Core: delete_all() now paginates through the vector store in batches of 1000 instead of listing once, so accounts with more memories than a single page (most vector stores default to ~100) had the remainder silently left behind (#6636)
Vector Stores: Cap Supabase search()/list() top_k at the vecs query limit of 1000 instead of erroring, and fix a col_info() crash by reading collection attributes directly instead of calling the removed describe() method (#6695)
Vector Stores: Set size on Elasticsearch KNN search queries, so results respect top_k instead of being capped at Elasticsearch’s default of 10 hits (#5910)

Changes:

Rerankers: LLMReranker’s default model is now gpt-5-mini (was gpt-4o-mini) (#6703)

v2.0.14

New Features:

Vector Stores: Add an Oracle AI Vector Search provider (oracledb) with connection pooling, HNSW/IVF indexes, JSON metadata filtering, and six selectable distance metrics (#5358)

Bug Fixes:

Vector Stores: Translate a "*" filter value in OpenSearch into an exists query for every key, not just identity keys. It was previously ignored or matched literally against the string "*", so a wildcard filter returned nothing (#6522)
Vector Stores: Re-raise errors from OpenSearch search() instead of returning [], so a transport, auth, or index misconfiguration surfaces instead of looking like zero matches. keyword_search() still degrades on failure, since it is a best-effort BM25 signal (#6519)
Vector Stores: Guard the text field in Milvus update() behind the _has_bm25_schema check, matching insert(), so updating a memory in a collection without the BM25 text/sparse schema no longer fails (#5705)

v2.0.13

Bug Fixes:

Vector Stores: Fix reset() silently leaving stale vectors behind on local (on-disk) Qdrant when the old collection directory could not be removed, for example an open file handle on Windows or NFS (#6412)
Core: Stop update() metadata from overwriting or injecting user_id, agent_id, run_id, or actor_id. These identity fields are immutable after creation, so passing them in metadata can no longer move a memory into a different tenant’s scope (#6278)
Vector Stores: Scope Pinecone delete_col()/reset() to the configured namespace instead of deleting the whole index, so resetting a namespaced Pinecone store no longer wipes out the other namespaces sharing that index (#6287)
Vector Stores: Convert Baidu Mochow’s raw L2 distance into a similarity score in search() (1 / (1 + distance)), so closer matches rank higher instead of lower, matching the Milvus provider and the rest of the VectorStoreBase contract (#6435)
LLMs: Read OPENAI_BASE_URL (was OPENAI_API_BASE) in OpenAIStructuredLLM, matching the official OpenAI SDK’s environment variable and the rest of the OpenAI-compatible providers (#6322)

Improvements:

LLMs: Remove a dead, no-op api_key attribute check from LLMBase.__init__ (#6460)

Changes:

Client: Remove the retrieval_criteria parameter from MemoryClient.update_project()/AsyncMemoryClient.update_project() and Project.update()/AsyncProject.update(). It was accepted and forwarded but never affected retrieval, so removing it is not a behavior change (#6313)

v2.0.12

New Features:

Memory (OSS): Accept text in Memory.update() and AsyncMemory.update(). data still works but is now deprecated, so prefer text in new code (#6044)

Bug Fixes:

Core: Coerce non-string entity IDs (user_id, agent_id, run_id) instead of crashing on .strip(), so passing an integer ID no longer raises AttributeError (#6206)
Core: Stop requiring langchain-core for the default async procedural memory path. The optional dependency is now only imported when you pass a custom LangChain LLM, matching the sync behavior (#6209)
Client: Encode dynamic URL path segments so IDs containing special characters no longer produce malformed requests (#5963)
LLMs: Skip temperature and top_p for newer Anthropic models that reject sampling parameters. Detection is automatic per model family and version, and the new enable_sampling_parameters config flag overrides it (#6211)
Vector Stores: Stop writing internal OutputData model fields as properties on Weaviate update() (#6149)
Vector Stores: Improve wildcard search handling in Milvus (#6187)
Vector Stores: Keep env-resolved Upstash Vector credentials after config validation. An env-var-only config previously passed validation and then failed to build (#5811)
Vector Stores: Restore the previous payload when a Neptune Analytics vector upsert fails inside update(), so a partial write can no longer leave the payload and embedding out of sync (#5824)

Changes:

LLMs: The Together default model is now MiniMaxAI/MiniMax-M3 (was mistralai/Mixtral-8x7B-Instruct-v0.1) (#6049)
LLMs: The xAI default model is now grok-4.3 (was grok-2-latest) (#6115)
Embeddings: The Together default embedding model is now intfloat/multilingual-e5-large-instruct at 1024 dimensions (was togethercomputer/m2-bert-80M-8k-retrieval at 768). If you use the Together embedder without pinning model, existing vectors were written at the old dimension: either re-embed them, or pin model and embedding_dims to the old values (#5989)
Rerankers: The Cohere default rerank model is now rerank-v3.5 (was rerank-english-v3.0) (#6055)

Security:

Vector Stores: Fix SQL and Cypher injection vulnerabilities in the PGVector, Azure MySQL, and Neptune providers (#4878)
Vector Stores: Validate Elasticsearch filter keys and values to prevent term query injection (#5980)
Dependencies: Require transformers>=5.3.0 to remediate GHSA-29pf-2h5f-8g72 (CVE-2026-4372) (#6110)

v2.0.11

Bug Fixes:

Embeddings: Guard against an embed_batch count mismatch in the OpenAI and Azure OpenAI embedders (#5966)
Memory: Re-raise LLM extraction failures instead of silently returning [] (#5878)
Vector Stores: Normalize vectors for the cosine distance strategy in FAISS (#5960)

Security:

Vector Stores: Validate OpenSearch filter values to prevent term query injection (#5986)
Vector Stores: Validate value types and escape quotes in Azure AI Search OData filters (#5983)
Vector Stores: Validate Databricks catalog/schema/table identifiers to prevent SQL injection (#5988)
Graph: Escape Neptune filter values in openCypher queries to prevent injection (#5982)

v2.0.10

New Features:

Client: Expose expiration_date on MemoryClient.update() and AsyncMemoryClient.update(): callers can now set or clear a memory’s expiration date; None is preserved and forwarded to the API (#5874)

Bug Fixes:

Memory (OSS): Apply remove_code_blocks() to the LangChain path in async _create_procedural_memory so code fences are stripped consistently (#5711)
Rerankers: Score HuggingFace cross-encoder results with per-document sigmoid instead of set-relative min-max, preventing a single low-score document from collapsing all relevance scores to zero (#5715)
Core: Validate and trim entity IDs (user_id, agent_id, run_id) in delete_all() for both sync and async Memory (#5735)
Vector Stores: Use .get() for hash and created_at in the Redis insert() and update() paths so entity payloads that omit those fields no longer raise KeyError (#5709)
Memory: Fix scale-threshold notices not firing for Redis and search-engine backends by resolving col_info() signature differences and adding num_docs to the count-extraction lookup (#5687)
Vector Stores: Escape special characters in Valkey FT.SEARCH tag filter values to prevent wildcard and operator injection through tenant-isolation filters (#5750)

v2.0.9

Bug Fixes:

Memory (OSS): Improve entity extraction precision by avoiding sentence-start common noun noise, preserving useful topic phrases, and exact-deduplicating entity links before semantic matching (#5829)

v2.0.8

New Features:

Embeddings: Add native embed_batch to five embedders for batched embedding requests: LM Studio, Together, HuggingFace, Vertex AI, and Google GenAI (#5609)

Bug Fixes:

Core: Guard against malformed image_url entries in parse_vision_messages to prevent crashes (#5631)
Core: Return attributed_to from get(), get_all(), and search() (#5629)
Core: Fix reset() only dropping the history table and leaving stale messages behind (#5541)
Core: Guard against an entity embed_batch count mismatch in the v3 add pipeline (#5604)
Core: Fix an async delete_all race condition that corrupted the entity store’s linked_memory_ids (#5553)
LLMs: Skip the JSON response_format for Groq compound models that reject it (#5513)
LLMs: Preserve reasoning fields during base-to-provider config conversion (#5638)
LLMs: Pass the configured anthropic_base_url to the Anthropic client (#5626)
LLMs: Stop the Azure provider from mutating and corrupting caller messages during content rewrite (#5731)
LLMs & Embeddings: Repair HTTP proxy support for httpx>=0.28 and preserve proxies in LlmFactory (#5447)
Embeddings: Forward embedding_dims to Titan V2 in the AWS Bedrock embedder (#5671)
Rerankers: Log reranking failures instead of swallowing them silently (#5717)
Rerankers: Clamp out-of-range LLM scores instead of mis-parsing them (#5635)
Rerankers: Export all five rerankers from the package root (#5636)
Vector Stores: Point the FastEmbed-missing warning at mem0ai[extras] (#5622)
Vector Stores: Preserve empty Azure AI Search update values (#5524)
Vector Stores: Add an auto_refresh option for OpenSearch Serverless compatibility (#3893)
Vector Stores: Wrap a scalar vector_id in a list for Chroma delete() (#5703)
Vector Stores: Wrap Chroma update() ids, embeddings, and metadatas in lists (#5757)
Vector Stores: Wrap a scalar vector_id in a list for Milvus delete() (#5704)
Vector Stores: Map all comparison operators in the Pinecone _create_filter() (#5707)
Vector Stores: Return None instead of {} from Chroma _generate_where_clause for empty filters (#5713)
Vector Stores: Return [[]] from the OpenSearch list() error path to honor the list() contract (#5727)
Vector Stores: Return [[]] from the Pinecone list() error path instead of a dict (#5706)
Vector Stores: Return [[]] for an uninitialized FAISS index to honor the list() contract (#5725)
Vector Stores: Wrap the MongoDB list() return in an outer list to match the interface contract (#5729)
Vector Stores: Deep-copy Redis DEFAULT_FIELDS so instances keep distinct dims (#5633)
Vector Stores: Pass the required vectors arg in Vertex AI list() and similarity search (#5627)
Vector Stores: Return None from Redis get() for missing IDs (#5625)
Vector Stores: Drop a stray print in Weaviate list_cols (#5637)
Graph: Keep distinct entities that share a substring prefix (#5630)
Client: Check the HTTP status before parsing the ping response in _validate_api_key (#5639)
Server: Fetch filtered dashboard memories beyond the default page (#5753)
Server: Return 404/400 instead of 502 for not-found and invalid input (#5634)
Server: Return 404 instead of 500 for a malformed API key id on revoke (#5640)
Server: Use 127.0.0.1 in the dashboard healthcheck to avoid IPv6 localhost resolution (#5612)

Improvements:

Vector Stores: Batch BM25 sparse encoding in Qdrant insert (#5592)

Security:

Vector Stores: Sanitize Milvus and Baidu filter values to prevent expression injection (#5746)
Vector Stores: Reject dict filter values in MongoDB to prevent NoSQL operator injection (#5748)

v2.0.7

New Features:

LLMs: Add Gemini via Vertex AI as LLM provider (#4030)
Embeddings: Add native embed_batch to OllamaEmbedding for batched embedding requests (#5415)

Bug Fixes:

Core: Fix api_error_handler silently dropping return values from async methods (#5540)
Core: Fix AsyncMemory.reset() not resetting the entity store (#5535)
Core: Fix async delete_all aborting on first error, leaving partial deletion (#5529)
Core: Skip messages without a content key in message parsers to prevent KeyError crashes (#5575)
Core: Preserve custom metadata fields during memory update (#5480)
LLMs: Fix Anthropic tool_choice format and tool response parsing (#5537)
LLMs: Fix Ollama json format mutating the caller’s messages list in-place (#5539)
LLMs: Omit None config values from Gemini GenerateContentConfig to prevent validation errors (#5528)
LLMs: Honor reasoning-model params in AzureOpenAIStructuredLLM (#5548)
LLMs: Honor reasoning-model params in OpenAIStructuredLLM (#5458)
LLMs: Send max_completion_tokens for the GPT-5 family across all providers (#5547)
LLMs: Accept and forward **kwargs in Together, LangChain, and Sarvam providers (#5556)
LLMs: Fix Bedrock AI21 response parse default using dict literal instead of set (#5527)
LLMs: Fix LiteLLM function-calling check blocking all calls on non-tool models (#5536)
LLMs: Fix HuggingFace provider using self.config instead of raw config parameter (#5538)
Embeddings: Honor aws_session_token in AWS Bedrock embeddings (#5566)
Rerankers: Respect config.top_k in Cohere and ZeroEntropy fallback paths (#5560)
Vector Stores: Fix FAISS filtered search dropping over-fetched candidates before filtering (#5453)
Vector Stores: Fix Weaviate reset() crashing with missing vector_size argument (#5531)
Vector Stores: Pass embedding dims in Weaviate reset() to avoid re-init crash (#5570)
Vector Stores: Fix MongoDB reset() passing wrong argument to create_col() (#5532)
Vector Stores: Fix Pinecone hybrid search crashing when filters is None (#5533)
Vector Stores: Fix Redis crashing on empty or None filters in search() and list() (#5446)
Vector Stores: Return None from get() for missing IDs in Milvus, Weaviate, and Supabase (#5562)
Vector Stores: Return None from ChromaDB get() for missing IDs (#5561)

v2.0.6

New Features:

Memory: Add a contextual OSS-to-Platform notices system that surfaces occasional, situation-aware messages (first run, scale/performance thresholds, slow queries, and when temporal/decay features are relevant) pointing to the corresponding Mem0 Platform capabilities; disable via MEM0_TELEMETRY=false (#5494)

Bug Fixes:

Memory: Prevent a crash in parse_vision_messages when vision support is disabled (#5487)
Vector Stores: Expose the https option on the Qdrant vector store configuration so TLS endpoints can be targeted explicitly (#5380)
Vector Stores: Use valid S3 Vectors entity index names, fixing index operations that failed on invalid names (#5416)
Vector Stores: Fix search() crashing with a TypeError in the LangChain vector store when a result score is None (#5072)
Vector Stores: Use is not None instead of a truthiness check for vector/payload in the PGVector update() path, so valid empty/zero values are no longer skipped (#5488)
Vector Stores: Index the Valkey memory field as TEXT rather than TAG so full-text search behaves correctly (#5443)
Vector Stores: Implement $not filter support in the ChromaDB vector store (#5485)

v2.0.5

New Features:

Memory: Warn at init time when hybrid/BM25 search silently degrades to semantic-only because the configured vector store does not implement keyword_search. Affected stores: Chroma, FAISS, Cassandra, LangChain, Neptune Analytics, S3 Vectors, Supabase, TurboPuffer, Valkey (#5444)
Memory: Add opt-in explain=True parameter to Memory.search() and AsyncMemory.search(). When enabled, each result includes a score_details dict with semantic_score, bm25_score, entity_boost, raw_score, max_possible_score, final_score, and threshold so callers can understand and tune retrieval ranking (#5102)

Bug Fixes:

Vector Stores: Normalize similarity scores to [0, 1] (higher = better) consistently across all backends. 11 adapters previously returned raw distance metrics (lower = better): FAISS, Chroma, Milvus, Redis, Cassandra, PGVector, S3 Vectors, Supabase, Valkey, Azure MySQL, and Vertex AI Vector Search: causing incorrect ranking in multi-store setups (#5391)
Memory: Parallelize entity boost searches in Memory.search() and AsyncMemory.search(). Previously up to 8 entities were embedded and queried sequentially (16 serial round-trips with remote embedders); all entity lookups now run concurrently, eliminating multi-second latency on entity-rich queries (#5377)
Memory: Reject empty or whitespace-only queries in Memory.search(), AsyncMemory.search(), MemoryClient.search(), and AsyncMemoryClient.search() before any embedding or API call is made. Also strips leading/trailing whitespace from valid queries (#5258)
LLMs: Add is_reasoning_model: Optional[bool] override to BaseLlmConfig (surfaced on OpenAILlmConfig and AzureOpenAILlmConfig). Fixes silent zero-extraction when using Azure deployments with versioned gpt-5.x names that the automatic name-based heuristic cannot recognize (#5327)
LLMs: Fix xAI LLM provider: add XAIConfig with xai_base_url, forward tools/tool_choice in generate_response(), and parse tool_calls in the response. Previously the provider raised AttributeError at init and silently dropped tool results (#5190)
Vector Stores: Fix PGVector ConnectionPool hang in Docker Compose environments where the app container starts before Postgres is DNS-resolvable: switched to open=False to avoid blocking constructor or silent zombie pool (#5155)
Vector Stores: Fix PGVector sslmode handling for PostgreSQL URIs: the sslmode query parameter is now correctly extracted and forwarded when building the async connection pool (#5308)
Vector Stores: Fix S3 Vectors list() not applying metadata filters: filtering is now done client-side after fetching, with pagination preserved and top_k applied after filtering to prevent pre-truncation of matching rows (#5018)
Vector Stores: Fix Upstash Vector search() routing all queries to the default namespace: namespace is now passed as a top-level keyword argument to query_many() instead of inside the per-query dict where it was silently ignored (#5202)
Core: Replace mutable default arguments with None sentinels in embedder configs and the proxy module, preventing cross-request state contamination (#5302)

v2.0.4

New Features:

Client: delete() and async delete() accept delete_linked (default False). When True, deleting a memory also removes the older memories it superseded (the v3 linked_memory_ids chain), transitively: the delete-side counterpart of latest_only, so a superseded memory does not resurface after the current one is deleted (#5270)

v2.0.3

Bug Fixes:

Vector Stores: PGVector adapter now supports rich filter operators (eq, ne, gt, gte, lt, lte, in, nin, contains, icontains, wildcard *, $or, $not) in search(), keyword_search(), and list(). Previously only exact-equality filters worked: operator dicts were silently stringified and returned zero results (#5263)
Server: Fixed /search endpoint returning 502 when user_id, agent_id, or run_id are sent as top-level request fields. The server now maps these into the filters dict before calling Memory.search(), matching the v3 API contract. Top-level entity ID fields are marked as deprecated in the OpenAPI schema and emit a warning log: clients should migrate to filters={"user_id": "..."} (#5263)

v2.0.2

Bug Fixes:

Telemetry: Stitch OSS and platform PostHog identities on MemoryClient init so $identify events fire and a single user is no longer tracked as two or three disconnected personas (#5040)
Security: Harden against SQL injection and prompt injection (#4997)

New Features:

SDK: Expose decay on project.update (#5062)

Improvements:

Plugin: Hand mem0 search decisions to the agent (#4992)

v2.0.1

Bug Fixes:

Client: Map user_id, agent_id, run_id entity params to filters in GET /memories (#4960)
Memory: Honor prompt param in vector store extraction pipeline (#4914)
Memory: Add missing text_lemmatized field in AsyncMemory._create_memory (#4886)
Memory: Merge same-key operator dicts in AND metadata filters (#4853)
LLMs: Narrow _is_reasoning_model check to not match gpt-5.x variants (#4746)
Vector Stores: Add ca_certs config option for Elasticsearch vector store (#3993)
Vector Stores: Add agent_id and run_id to Elasticsearch/OpenSearch default mappings (#4906)
Embeddings: Set FastEmbed embedding_dims from model metadata at init (#4711)

Security:

Bump vulnerable dependencies to patched versions (#4835)

v2.0.0

Major Release: Python SDK with V3 memory pipeline, ADD-only extraction, and cleaned-up API surface.New Features:

Single-Pass Extraction: Replaced 2-LLM-call pipeline with additive extraction using ADDITIVE_EXTRACTION_PROMPT. Memories accumulate via linked_memory_ids: no more UPDATE/DELETE events (#4805)
Hybrid Search: Combined semantic + BM25 keyword matching + entity boost with additive scoring. Native keyword_search() added to 15 vector store adapters (Qdrant, Elasticsearch, OpenSearch, Azure AI Search, Weaviate, Redis, PGVector, Pinecone, Databricks, MongoDB, Milvus, Baidu, Upstash, Azure MySQL, Vertex AI) (#4805)
Entity Extraction & Linking: spaCy-based entity extraction with second vector collection ({collection}_entities) for cross-memory relationship retrieval. Optional dependency: pip install mem0ai[nlp] (#4805)
Batch Operations: Batch embedding, batch persist, and batch entity linking (8-phase pipeline) for both sync Memory and async AsyncMemory at full parity (#4805)
Message Persistence: SQLite-based rolling window (10 messages per session scope) for LLM context (#4805)
Valkey Cluster Mode: Added cluster_mode parameter for Valkey Cluster Mode Enabled (CME) deployments (#4759)
V3 API Endpoints: MemoryClient.add() now posts to /v3/memories/add/; MemoryClient.get_all() posts to /v3/memories/ and returns a paginated envelope {"count": int, "next": str | None, "previous": str | None, "results": [...]} (#4856)
Default model: gpt-5-mini is now the default across OpenAILLM, OpenAIStructuredLLM, AzureOpenAILLM, AzureOpenAIStructuredLLM, and LiteLLM fallback (#4829)

Breaking Changes:

add() returns ADD-only events: No more "UPDATE" or "DELETE" events. Memories accumulate; nothing is overwritten (#4805)
search() default threshold is now 0.1: Pass threshold=0.0 for previous behavior (#4805)
search() score is now a combined multi-signal score: The top-level score fuses semantic similarity, BM25 keyword match, entity signals, and temporal boosts into one value. Absolute numbers shift versus the old raw cosine score; retune any hard thresholds against representative queries (#4805, #4836)
search() default rerank is now False: Pass rerank=True for previous behavior (#4805)
top_k default changed 100 → 20 in Memory.get_all() and Memory.search() (sync + async). Pass top_k=100 explicitly to restore the old behavior (#4843)
Entity ID validation: user_id / agent_id / run_id are trimmed; empty-string and whitespace-only values now raise ValueError (#4843)
Search params validation: threshold must be a number in [0, 1]; top_k must be a non-negative integer: invalid inputs raise ValueError (#4843)
messages in Memory.add() rejects invalid types: Passing None or non-(str | dict | list) values raises Mem0ValidationError (error_code="VALIDATION_003") (#4843)
qdrant-client>=1.12.0 required: Upgrade from >=1.9.1 (#4805)
org_id and project_id removed: Removed from MemoryClient constructor and all method signatures (#4740)
External Graph Store Removed (OSS): mem0/memory/graph_memory.py, memgraph_memory.py, kuzu_memory.py, apache_age_memory.py, and mem0/graphs/ (Neo4j / Memgraph / Kuzu / Apache AGE / Neptune drivers) deleted, about 4,000 lines. The external graph store integration is no longer part of the OSS SDK; graph drivers (neo4j, memgraph, kuzu, etc.) can be uninstalled. Graph memory now runs natively as built-in entity linking. Remove enable_graph and graph_store from your config (#4805)
enable_graph removed from Client SDK: Graph memory now runs automatically and no longer needs a flag. Remove enable_graph from MemoryClient.add() / search() / get_all() / update_project() calls (#4776)
custom_fact_extraction_prompt renamed to custom_instructions: Update config and memory module references (#4740)
Typed option classes: Added Pydantic v2 typed classes: AddMemoryOptions, SearchMemoryOptions, GetAllMemoryOptions, DeleteAllMemoryOptions, UpdateMemoryOptions, ProjectUpdateOptions (#4740)

Security:

FAISS: Prevent arbitrary code execution via pickle deserialization in FAISS vector store (#4833)

Bug Fixes:

V3 migration crashes: Fixed crashes in the v3 migration path; entity linking on OSS is now functional across Qdrant and Milvus backends (#4836)
Qdrant entity store: Entity store now shares the existing Qdrant client when using embedded mode (path=...), eliminating RocksDB lock contention between the main and entity collections (#4836)
Reranker: Fixed incorrect use of SentenceTransformer for cross-encoder reranker models: switched to CrossEncoder API for proper scoring (#4806)
S3 Vectors: Handle vector=None in update() to prevent boto3 validation error when event=NONE (#4594)
LLMs: Made OpenAI store parameter opt-in to prevent leaking to non-OpenAI backends like Google Gemini (#4757)
LLMs: Forward response_format to Azure OpenAI API to prevent JSON parsing failures (#4689)
Core: Guard temp_uuid_mapping lookups against LLM-hallucinated IDs with safe .get() and warnings (#4674)
Client: Prevent MemoryClient.feedback() telemetry TypeError by merging feedback data into single payload (#4795)

Improvements:

Telemetry: Sample OSS hot-path events at 10% via PostHog before_send hook to reduce event volume (#4771)

See the OSS v2 to v3 migration guide and Platform migration guide for upgrade instructions.

v1.0.11

New Features & Updates:

SDK: Added multilingual parameter to project update (#4314)

Bug Fixes:

LLMs: Fixed Groq model configuration (#4700)
Core: Prevented thread and memory leaks from PostHog telemetry (#4535)
Vector Stores: Used DatetimeRange for datetime string values in Qdrant range filters (#4659)
Configs: Added missing ConfigDict to vector store configs (Elasticsearch, MongoDB, Neptune, OpenSearch, PGVector, Supabase, Valkey) (#4656)

v1.0.10

New Features & Updates:

LLMs: Added MiniMax provider support for AWS Bedrock (#4609)

Bug Fixes:

Configs: Migrated CassandraConfig and AzureMySQLConfig to pydantic v2 ConfigDict (#4646)
LLMs: Forward response_format to OpenAI-compatible API for DeepSeek (#4635)
LLMs: Forward response_format to OpenAI-compatible API for vLLM (#4608)
Vector Stores: Only list authorized collections when listing MongoDB collections (#3888)
Core: Reset graph database in Memory.reset() (#4185)
Core: Make AsyncMemory.from_config a regular classmethod (#4183)

v1.0.9

New Features & Updates:

LLMs: Added reasoning_effort parameter support for reasoning models (#4461)

Bug Fixes:

Core: Preserved original actor_id during memory update (#4570)
Core: Set updated_at on creation and preserve pre-existing created_at (#4499)
Core: Centralized entity cleanup and skip malformed LLM relation dicts (#4515)
Core: Removed README.md from wheel shared-data (#4052)
Vector Stores: Handled vector=None in Milvus and Qdrant update methods (#4568)
Vector Stores: Rebuilt FAISS index on vector deletion (#4178)

Improvements:

Embeddings: Updated default Gemini and Vertex AI embedder model to gemini-embedding-001 (#4571)

v1.0.8

New Features & Updates:

Vector Stores: Integrated Turbopuffer as a vector database provider (#4428)
LLMs: Added MiniMax LLM provider (#4431)

Bug Fixes:

Core: Fixed merging of multiple filter operators for the same key (#4559)
Core: Prevented in-place mutation of metadata in _create_memory (#4529)
Core: Preserved custom metadata when updating memory (#4495)
Core: Handled chatty LLM responses in JSON parsing (#4525)
Core: Prevented double embedding in mem0.add (#3996)
Core: Raised ValueError when deleting nonexistent memory (#4455)
Core: Cleaned up graph store data on Memory.delete() (#4505)
Vector Stores: Prevented SQL injection in Databricks vector store (#4558)
Vector Stores: Upgraded MongoDB vector store from deprecated knnVector to GA vectorSearch (#3995)
Vector Stores: Prevented embedding corruption in Valkey and Redis when vector is None (#4362)
Vector Stores: Accepted default /tmp/chroma path in ChromaDbConfig validator (#4179)
Vector Stores: Wrapped vector and payload in lists for Langchain.update (#4446)
Graph: Soft-delete graph relationships instead of hard DELETE (#4188)
Graph: Sanitized hyphens in Neo4j Cypher relationship names (#4154)
Graph: Used root LLM config as fallback for graph store instead of hardcoded OpenAI default (#4466)
Qdrant: Fixed do not remove local path on init (#4475)
Qdrant: Implemented enhanced metadata filtering operators (#4127)
Embeddings: Fixed OpenAI embedding dimensions (#4481)
LLMs: Omitted topP for Anthropic Converse in Bedrock; used AWSBedrockConfig in LlmFactory (#4469)
LLMs: Avoided sending both temperature and top_p to Anthropic API (#4471)
LLMs: Handled None content and empty candidates in GeminiLLM parsing (#4462)
LLMs: Added missing _parse_response to AzureOpenAIStructuredLLM (#4434)
History: Added timestamps for DELETE operations in history (#4492)

Improvements:

Vector Stores: Added vector validation to OpenSearchDB to ensure non-null, non-empty, and correct-dimension vectors (#4472)

v1.0.7

Bug Fixes:

Core: Fixed control characters in LLM JSON responses causing parse failures (#4420)
Core: Replaced hardcoded US/Pacific timezone references with timezone.utc (#4404)
Core: Preserved http_auth in _safe_deepcopy_config for OpenSearch (#4418)
Core: Normalized malformed LLM fact output before embedding (#4224)
Embeddings: Pass encoding_format='float' in OpenAI embeddings for proxy compatibility (#4058)
LLMs: Fixed Ollama to pass tools to client.chat and parse tool_calls from response (#4176)
Reranker: Support nested LLM config in LLMReranker for non-OpenAI providers (#4405)
Vector Stores: Cast vector_distance to float in Redis search (#4377)

Improvements:

Embeddings: Improved Ollama embedder with model name normalization and error handling (#4403)

v1.0.6

Bug Fixes:

Telemetry: Fixed telemetry vector store initialization still running when MEM0_TELEMETRY is disabled (#4351)
Core: Removed destructive vector_store.reset() call from delete_all() that was wiping the entire vector store instead of deleting only the target memories (#4349)
OSS: OllamaLLM now respects the configured URL instead of always falling back to localhost (#4320)
Core: Fixed KeyError when LLM omits the entities key in tool call response (#4313)
Prompts: Ensured JSON instruction is included in prompts when using json_object response format (#4271)
Core: Fixed incorrect database parameter handling (#3913)

Dependencies:

Updated LangChain dependencies to v1.0.0 (#4353)
Bumped protobuf dependency to 5.29.6 and extended upper bound to <7.0.0 (#4326)

v1.0.5

Telemetry Fix
- Fixed an issue where the PostHog client was initialized even after telemetry was disabled. Although events were not captured, the client was unnecessarily initialized.

v1.0.4

New Features & Updates:

Memory Update:
- Added timestamp parameter to update(): accepts Unix epoch (int/float) or ISO 8601 string

v1.0.3

New Features & Updates:

Project Settings:
- Added inclusion prompt, exclusion prompt, memory depth, and usecase setting

v1.0.2

New Features & Updates:

Vector Stores:
- Added DriverInfo metadata to MongoDB vector store

v1.0.1

New Features & Updates:

Vector Stores:
- Added Apache Cassandra vector store support
Embeddings:
- Added FastEmbed embedding support for local embeddings
Graph Store:
- Added configurable embedding similarity threshold for graph store node matching

Bug Fixes:

Core:
- Fixed condition check for memories_result type in Memory class
- Fixed list_memories endpoint Pydantic validation error
- Fixed memory deletion not removing from vector store

v1.0.0

New Features & Updates:

Vector Stores:
- Added Azure MySQL support
- Added Azure AI Search Vector Store support
LLMs:
- Added Tool Call support for LangchainLLM
- Enabled custom model and parameters for Hugging Face with huggingface_base_url
- Updated default LLM configuration
Rerankers:
- Added reranker support: Cohere, ZeroEntropy, Hugging Face, Sentence Transformers, and LLMs
Core:
- Added metadata filtering for OSS
- Added Assistant memory retrieval
- Enabled async mode as default

Improvements:

Prompts:
- Improved prompt for better memory retrieval
Dependencies:
- Updated dependency compatibility with OpenAI 2.x
Validation:
- Validated embedding_dims for Kuzu integration

Bug Fixes:

Vector Stores:
- Fixed Databricks Vector Store integration
- Fixed Milvus DB bug and added test coverage
- Fixed Weaviate search method
LLMs:
- Fixed bug with thinking LLM in vLLM

v0.1.118

New Features & Updates:

Vector Stores:
- Added Valkey vector store support
- Added support for ChromaDB Cloud
- Added Mem0 vector store backend integration for Neptune Analytics
Graph Store:
- Added Neptune-DB graph store with vector store
Core:
- Implemented structured exception classes with error codes and suggested actions

Improvements:

Dependencies:
- Updated OpenAI dependency and improved Ollama compatibility
Testing:
- Added Weaviate DB test
- Added comprehensive test suite for SQLiteManager
Documentation:
- Updated category docs
- Updated Search V2 / Get All V2 filters documentation
- Refactored AWS example title
- Fixed Quickstart cURL example

Bug Fixes:

Vector Stores:
- Databricks bug fixes
- Fixed S3 Vectors memory initialization issue from configuration
Core:
- Fixed JSON parsing with new memories
- Replaced hardcoded LLM provider with provider from configuration
LLMs:
- Fixed Bedrock Anthropic models to use system field

v0.1.117

New Features & Updates:

OpenMemory:
- Added memory export / import feature
- Added vector store integrations: Weaviate, FAISS, PGVector, Chroma, Redis, Elasticsearch, Milvus
- Added export_openmemory.sh migration script
Vector Stores:
- Added Amazon S3 Vectors support
- Added Databricks Mosaic AI vector store support
- Added support for OpenAI Store
Graph Memory: Added support for graph memory using Kuzu
Azure: Added Azure Identity for Azure OpenAI and Azure AI Search authentication
Elasticsearch: Added headers configuration support

Improvements:

Added custom connection client to enable connecting to local containers for Weaviate
Updated configuration AWS Bedrock
Fixed dependency issues and tests; updated docstrings
Documentation:
- Fixed Graph Docs page missing in sidebar
- Updated integration documentation
- Added version param in Search V2 API documentation
- Updated Databricks documentation and refactored docs
- Updated favicon logo
- Fixed typos and Typescript docs

Bug Fixes:

Baidu: Added missing provider for Baidu vector DB
MongoDB: Replaced query_vector args in search method
Fixed new memory mistaken for current
AsyncMemory._add_to_vector_store: handled edge case when no facts found
Fixed missing commas in Kuzu graph INSERT queries
Fixed inconsistent created and updated properties for Graph
Fixed missing app_id on client for Neptune Analytics
Correctly pick AWS region from environment variable
Fixed Ollama model existence check

Refactoring:

PGVector: Use internal connection pools and context managers

v0.1.116

New Features & Updates:

Pinecone: Added namespace support and improved type safety
Milvus: Added db_name field to MilvusDBConfig
Vector Stores: Added multi-id filters support
Vercel AI SDK: Migration to AI SDK V5.0
Python Support: Added Python 3.12 support
Graph Memory: Added sanitizer methods for nodes and relationships
LLM Monitoring: Added monitoring callback support

Improvements:

Performance:
- Improved async handling in AsyncMemory class
Documentation:
- Added async add announcement
- Added personalized search docs
- Added Neptune examples
- Added V5 migration docs
Configuration:
- Refactored base class config for LLMs
- Added sslmode for pgvector
Dependencies:
- Updated psycopg to version 3
- Updated Docker compose

Bug Fixes:

Tests:
- Fixed failing tests
- Restricted package versions
Memgraph:
- Fixed async attribute errors
- Fixed n_embeddings usage
- Fixed indexing issues
Vector Stores:
- Fixed Qdrant cloud indexing
- Fixed Neo4j Cypher syntax
- Fixed LLM parameters
Graph Store:
- Fixed LM config prioritization
Dependencies:
- Fixed JSON import for psycopg

Refactoring:

Google AI: Refactored from Gemini to Google AI
Base Classes: Refactored LLM base class configuration

v0.1.115

New Features & Updates:

Enhanced project management via client.project and AsyncMemoryClient.project interfaces
Full support for project CRUD operations (create, read, update, delete)
Project member management: add, update, remove, and list members
Manage project settings including custom instructions, categories, retrieval criteria, and graph enablement
Both sync and async support for all project management operations

Improvements:

Documentation:
- Added detailed API reference and usage examples for new project management methods.
- Updated all docs to use client.project.get() and client.project.update() instead of deprecated methods.
Deprecation:
- Marked get_project() and update_project() as deprecated (these methods were already present); added warnings to guide users to the new API.

Bug Fixes:

Tests:
- Fixed Gemini embedder and LLM test mocks for correct error handling and argument structure.
vLLM:
- Fixed duplicate import in vLLM module.

v0.1.114

New Features:

OpenAI Agents: Added OpenAI agents SDK support
Amazon Neptune: Added Amazon Neptune Analytics graph_store configuration and integration
vLLM: Added vLLM support

Improvements:

Documentation:
- Added SOC2 and HIPAA compliance documentation
- Enhanced group chat feature documentation for platform
- Added Google AI ADK Integration documentation
- Fixed documentation images and links
Setup: Fixed Mem0 setup, logging, and documentation issues

Bug Fixes:

MongoDB: Fixed MongoDB Vector Store misaligned strings and classes
vLLM: Fixed missing OpenAI import in vLLM module and call errors
Dependencies: Fixed CI issues related to missing dependencies
Installation: Reverted pip install changes

v0.1.113

Bug Fixes:

Gemini: Fixed Gemini embedder configuration

v0.1.112

New Features:

Memory: Added immutable parameter to add method
OpenMemory: Added async_mode parameter support

Improvements:

Documentation:
- Enhanced platform feature documentation
- Fixed documentation links
- Added async_mode documentation
MongoDB: Fixed MongoDB configuration name

Bug Fixes:

Bedrock: Fixed Bedrock LLM, embeddings, tools, and temporary credentials
Memory: Fixed memory categorization by updating dependencies and correcting API usage
Gemini: Fixed Gemini Embeddings and LLM issues

v0.1.111

New Features:

OpenMemory:
- Added OpenMemory augment support
- Added OpenMemory Local Support using new library
vLLM: Added vLLM support integration

Improvements:

Documentation:
- Added MCP Client Integration Guide and updated installation commands
- Improved Agent Id documentation for Mem0 OSS Graph Memory
Core: Added JSON parsing to solve hallucination errors

Bug Fixes:

Gemini: Fixed Gemini Embeddings migration

v0.1.110

New Features:

Baidu: Added Baidu vector database integration

Improvements:

Documentation:
- Updated changelog
- Fixed example in quickstart page
- Updated client.update() method documentation in OpenAPI specification
OpenSearch: Updated logger warning

Bug Fixes:

CI: Fixed failing CI pipeline

v0.1.109

New Features:

AgentOps: Added AgentOps integration
LM Studio: Added response_format parameter for LM Studio configuration
Examples: Added Memory agent powered by voice (Cartesia + Agno)

Improvements:

AI SDK: Added output_format parameter
Client: Enhanced update method to support metadata
Google: Added Google Genai library support

Bug Fixes:

Build: Fixed Build CI failure
Pinecone: Fixed pinecone for async memory

v0.1.108

New Features:

MongoDB: Added MongoDB Vector Store support
Client: Added client support for summary functionality

Improvements:

Pinecone: Fixed pinecone version issues
OpenSearch: Added logger support
Testing: Added python version test environments

v0.1.107

Improvements:

Documentation:
- Updated Livekit documentation migration
- Updated OpenMemory hosted version documentation
Core: Updated categorization flow
Storage: Fixed migration issues

v0.1.106

New Features:

Cloudflare: Added Cloudflare vector store support
Search: Added threshold parameter to search functionality
API: Added wildcard character support for v2 Memory APIs

Improvements:

Documentation: Updated README docs for OpenMemory environment setup
Core: Added support for unique user IDs

Bug Fixes:

Core: Fixed error handling exceptions

v0.1.104

Bug Fixes:

Vector Stores: Fixed GET_ALL functionality for FAISS and OpenSearch

v0.1.103

New Features:

LLM: Added support for OpenAI compatible LLM providers with baseUrl configuration

Improvements:

Documentation:
- Fixed broken links
- Improved Graph Memory features documentation clarity
- Updated enable_graph documentation
TypeScript SDK: Updated Google SDK peer dependency version
Client: Added async mode parameter

v0.1.102

New Features:

Examples: Added Neo4j example
AI SDK: Added Google provider support
OpenMemory: Added LLM and Embedding Providers support

Improvements:

Documentation:
- Updated memory export documentation
- Enhanced role-based memory attribution rules documentation
- Updated API reference and messages documentation
- Added Mastra and Raycast documentation
- Added NOT filter documentation for Search and GetAll V2
- Announced Claude 4 support
Core:
- Removed support for passing string as input in client.add()
- Added support for sarvam-m model
TypeScript SDK: Fixed types from message interface

Bug Fixes:

Memory: Prevented saving prompt artifacts as memory when no new facts are present
OpenMemory: Fixed typos in MCP tool description

v0.1.101

New Features:

Neo4j: Added base label configuration support

Improvements:

Documentation:
- Updated Healthcare example index
- Enhanced collaborative task agent documentation clarity
- Added criteria-based filtering documentation
OpenMemory: Added cURL command for easy installation
Build: Migrated to Hatch build system

v0.1.100

New Features:

Memory: Added Group Chat Memory Feature support
Examples: Added Healthcare assistant using Mem0 and Google ADK

Bug Fixes:

SSE: Fixed SSE connection issues
MCP: Fixed memories not appearing in MCP clients added from Dashboard

v0.1.99

New Features:

OpenMemory: Added OpenMemory support
Neo4j: Added weights to Neo4j model
AWS: Added support for OpenSearch Serverless
Examples: Added ElizaOS Example

Improvements:

Documentation: Updated Azure AI documentation
AI SDK: Added missing parameters and updated demo application
OSS: Fixed AOSS and AWS BedRock LLM

v0.1.98

New Features:

Neo4j: Added support for Neo4j database
AWS: Added support for AWS Bedrock Embeddings

Improvements:

Client: Updated delete_users() to use V2 API endpoints
Documentation: Updated timestamp and dual-identity memory management docs
Neo4j: Improved Neo4j queries and removed warnings
AI SDK: Added support for graceful failure when services are down

Bug Fixes:

Fixed AI SDK filters
Fixed new memories wrong type
Fixed duplicated metadata issue while adding/updating memories

v0.1.97

New Features:

HuggingFace: Added support for HF Inference

Bug Fixes:

Fixed proxy for Mem0

v0.1.96

New Features:

Vercel AI SDK: Added Graph Memory support

Improvements:

Documentation: Fixed timestamp and README links
Client: Updated TS client to use proper types for deleteUsers
Dependencies: Removed unnecessary dependencies from base package

v0.1.95

Improvements:

Client: Fixed Ping Method for using default org_id and project_id
Documentation: Updated documentation

Bug Fixes:

Fixed mem0-migrations issue

v0.1.94

New Features:

Integrations: Added Memgraph integration
Memory: Added timestamp support
Vector Stores: Added reset function for VectorDBs

Improvements:

Documentation:
- Updated timestamp and expiration_date documentation
- Fixed v2 search documentation
- Added “memory” in EC “Custom config” section
- Fixed typos in the json config sample

v0.1.93

Improvements:

Vector Stores: Initialized embedding_model_dims in all vectordbs

Bug Fixes:

Documentation: Fixed agno link

v0.1.92

New Features:

Memory: Added Memory Reset functionality
Client: Added support for Custom Instructions
Examples: Added Fitness Checker powered by memory

Improvements:

Core: Updated capture_event
Documentation: Fixed curl for v2 get_all

Bug Fixes:

Vector Store: Fixed user_id functionality
Client: Various client improvements

v0.1.91

New Features:

LLM Integrations: Added Azure OpenAI Embedding Model
Examples:
- Added movie recommendation using grok3
- Added Voice Assistant using Elevenlabs

Improvements:

Documentation:
- Added keywords AI
- Reformatted navbar page URLs
- Updated changelog
- Updated openai.mdx
FAISS: Silenced FAISS info logs

v0.1.90

New Features:

LLM Integrations: Added Mistral AI as LLM provider

Improvements:

Documentation:
- Updated changelog
- Fixed memory exclusion example
- Updated xAI documentation
- Updated YouTube Chrome extension example documentation

Bug Fixes:

Core: Fixed EmbedderFactory.create() in GraphMemory
Azure OpenAI: Added patch to fix Azure OpenAI
Telemetry: Fixed telemetry issue

v0.1.89

New Features:

Langchain Integration: Added support for Langchain VectorStores
Examples:
- Added personal assistant example
- Added personal study buddy example
- Added YouTube assistant Chrome extension example
- Added agno example
- Updated OpenAI Responses API examples
Vector Store: Added capability to store user_id in vector database
Async Memory: Added async support for OSS

Improvements:

Documentation: Updated formatting and examples

v0.1.87

New Features:

Upstash Vector: Added support for Upstash Vector store

Improvements:

Code Quality: Removed redundant code lines
Build: Updated MAKEFILE
Documentation: Updated memory export documentation

v0.1.86

Improvements:

FAISS: Added embedding_dims parameter to FAISS vector store

v0.1.84

New Features:

Langchain Embedder: Added Langchain embedder integration

Improvements:

Langchain LLM: Updated Langchain LLM integration to directly pass the Langchain object LLM

v0.1.83

Bug Fixes:

Langchain LLM: Fixed issues with Langchain LLM integration

v0.1.82

New Features:

LLM Integrations: Added support for Langchain LLMs, Google as new LLM and embedder
Development: Added development docker compose

Improvements:

Output Format: Set output_format=‘v1.1’ and updated documentation

Documentation:

Integrations: Added LMStudio and Together.ai documentation
API Reference: Updated output_format documentation
Integrations: Added PipeCat integration documentation
Integrations: Added Flowise integration documentation for Mem0 memory setup

Bug Fixes:

Tests: Fixed failing unit tests

v0.1.79

New Features:

FAISS Support: Added FAISS vector store support

v0.1.78

New Features:

Livekit Integration: Added Mem0 livekit example
Evaluation: Added evaluation framework and tools

Documentation:

Multimodal: Updated multimodal documentation
Examples: Added examples for email processing
API Reference: Updated API reference section
Elevenlabs: Added Elevenlabs integration example

Bug Fixes:

OpenAI Environment Variables: Fixed issues with OpenAI environment variables
Deployment Errors: Added package.json file to fix deployment errors
Tools: Fixed tools issues and improved formatting
Docs: Updated API reference section for expiration date

v0.1.77

Bug Fixes:

OpenAI Environment Variables: Fixed issues with OpenAI environment variables
Deployment Errors: Added package.json file to fix deployment errors
Tools: Fixed tools issues and improved formatting
Docs: Updated API reference section for expiration date

v0.1.76

New Features:

Supabase Vector Store: Added support for Supabase Vector Store
Supabase History DB: Added Supabase History DB to run Mem0 OSS on Serverless
Feedback Method: Added feedback method to client

Bug Fixes:

Azure OpenAI: Fixed issues with Azure OpenAI
Azure AI Search: Fixed test cases for Azure AI Search

v3.3.1

Bug Fixes:

Config (OSS): ConfigManager no longer injects OpenAI’s default baseURL and model into non-OpenAI LLM providers. The defaults now apply only to openai and openai_structured, so a DeepSeek, xAI, or other provider config without an explicit baseURL or model falls back to that provider’s own defaults and env vars (DEEPSEEK_API_BASE, XAI_API_BASE, …) instead of pointing at OpenAI with an OpenAI model name (#7350)
Vector Stores: Turbopuffer filters now apply every operator (eq, ne, gt, gte, lt, lte, in, nin). Only the range operators were read before, so eq, ne, in, and nin were silently dropped. A "*" value now matches anything instead of nothing, an array value is treated as in, and an unsupported operator throws (#6578)
Vector Stores: Turbopuffer search() with euclidean_squared now maps the unbounded squared distance to 1 / (1 + distance) instead of 1 - distance, which went negative and inverted ranking for any distance above 1 (#6580)
Packaging: pg, @types/pg, and natural are now optional peer dependencies, and pg / @types/pg accept caret ranges instead of the exact 8.11.3 / 8.11.0 pins, so installing mem0ai no longer pulls in pg or conflicts with an app’s own pg version. The PGVector store now imports pg only when it is used, so install it alongside mem0ai (npm install pg) if you use that store. @types/jest moved from peer dependencies to dev dependencies (#7450)

v3.3.0

New Features:

Client: Add User Profiles to MemoryClient: getProfile(), generateProfile(), getProfileSettings(), updateProfileSettings(), sampleProfiles(), and getProfileJob(). A profile is a structured, always-current JSON summary of one user, shaped by a JSON Schema you configure per project and filled by an LLM from that user’s memories. Generation is asynchronous. Every job POST carries an Idempotency-Key; to retry a lost request without starting a second job, pass the same idempotencyKey on each attempt (#7340)

v3.2.0

Improvements:

Client: Requests now carry X-Mem0-Source, X-Application and X-Mem0-Client, matching the Python SDK. The first two are set-once so an outer wrapper keeps its identity; the third is append-only and reports the whole layer chain. Read from MEM0_SOURCE, MEM0_APPLICATION and MEM0_CLIENT_STACK when set (#7326)
Client: The SDK version in X-Mem0-Client is injected at build time rather than hardcoded, so it cannot go stale at the next release (#7326)

v3.1.8

Improvements:

OSS notices: Notice configuration now comes from a static, cacheable repository file with a bundled disabled fallback and deterministic rollout assignment, instead of calling PostHog’s feature-flag evaluation API. This keeps notices fail-safe when the remote config is unavailable and removes the PostHog feature-flag request from notice evaluation (#7185)

v3.1.7

Bug Fixes:

Vector Stores: Redis and Valkey search() / get() / list() now preserve agent_id, run_id, and user_id as snake_case in the returned payload. The shared payload formatter camelCased every key including those three identity fields, so entity ids came back as agentId / runId / userId, inconsistent with every other vector store (#6902)
Memory (OSS): Embedding-cache lookups now use Object.prototype.hasOwnProperty.call() instead of the in operator or a falsy || check. Memory text matching an inherited Object.prototype property name (constructor, toString, and similar) previously short-circuited the lookup and resolved to that inherited value instead of computing a real embedding, silently corrupting the stored vector (#6903)
Client: getAll() now sends page and pageSize as independent query params instead of requiring both to be set before either was sent. Passing only pageSize without page previously had it silently dropped, so results came back at the server’s default page size (#6900)
LLMs: Add providerOverride to the Bedrock LLMConfig, an explicit provider name for when model is an application inference profile ARN whose opaque ID has no provider substring for extractProvider() to detect. Without it, those ARNs threw before the provider-specific settings could be initialized (#6899)

Security:

Dependencies: Resolved 17 additional high and critical severity dependency vulnerabilities across 5 pnpm workspaces (mem0-ts, vercel-ai-sdk, n8n-nodes-mem0, zapier-mem0, server/dashboard) via pnpm.overrides and a tar patch (#7032)

v3.1.6

New Features:

Vector Stores: Add an Oracle AI Vector Search vector store (oracledb) to the OSS SDK, with pooled connections, HNSW/IVF indexes, an optional indexAccuracy target, JSON payload filtering, and six selectable distance metrics (#6690)

Bug Fixes:

Core: Percent-escape %, &, and = in userId, agentId, and runId when building the session scope key for the recent-conversation buffer, so the key stays unambiguous for ids containing those characters. Ordinary ids keep their existing key; an id already containing %, &, or = maps to a new key, so its buffer starts empty once and refills on the next add(). Stored memories are unaffected (#6892)
Vector Stores: Close and clear the Oracle pool that Mem0 created when initialize() fails, and reset the cached init promise so the next call retries instead of replaying the rejection forever. A caller-supplied client is left untouched (#6839)
Vector Stores: Validate Oracle insert() batches before touching the database: ids and payloads must match vectors in length, so a short array can no longer write rows with undefined ids or silently drop payloads (#6839)

Improvements:

Vector Stores: Cut Oracle round trips. insert() now sends the whole batch through one executeMany() with explicit bindDefs instead of one INSERT per vector, list() reads the total from a COUNT(*) OVER () window in the same statement instead of issuing a second count query, and an unfiltered search() adds the VECTOR_INDEX_TRANSFORM hint so the vector index is used (#6835)

Security:

Dependencies: Patched 8 high and 18 medium severity dependency vulnerabilities across the pnpm workspace via pnpm.overrides (undici, brace-expansion, ip-address) (#6847)

v3.1.5

New Features:

Client: Add agentCustomInstructions to PromptUpdatePayload, AddMemoryOptions, and ProjectResponse. It sets a second extraction instruction set that applies only to agent-scoped memories: an add passing agentId without userId uses it, one passing both splits by attribution, and while it is unset customInstructions continues to apply to every memory (#6809)

v3.1.4

New Features:

Client: Add referenceDate and keywordSearch to SearchMemoryOptions, keeping the TypeScript client’s typed search options in sync with the Platform API (#6696)

Bug Fixes:

Client: Take the /v1/ping/ identity/telemetry call off the request critical path. Every method previously did if (this.telemetryId === "") await this.ping();, blocking the first call on each client instance on a network round trip; the ping now resolves in the background through a shared, credentials-keyed promise cache (FIFO-capped at 50 entries), so concurrent clients on the same host/API key share one ping instead of issuing one each, and a failed ping is not cached so the process can retry (#6788)
Client: deleteUsers() now issues its per-entity DELETE requests through the shared fetch()-based helper instead of the axios instance, which defaulted to keepAlive: false. Deleting every user, agent, app, and run now reuses pooled connections instead of paying a fresh TCP/TLS handshake per entity (#6788)
Memory (OSS): Stop add() metadata from setting or overwriting a memory’s identity scope. Metadata now runs through the same stripIdentityKeys() helper as update(), so user_id/agent_id/run_id (snake_case or camelCase) and actor_id passed in metadata can no longer place a memory into a scope the caller didn’t request through userId/agentId/runId/filters (#6377)
Vector Stores: Fix Redis search() and list() building an invalid, empty RediSearch filter expression when filters was an empty object or contained only null/undefined values. The shared buildRedisFilterExpr() helper now falls back to "*" (match all) instead of joining zero conditions into "" (#6014)

v3.1.3

New Features:

Vector Stores: Add Qdrant server-side BM25 keywordSearch() (requires Qdrant >= 1.15.2) plus payload filter indexes, so keyword search runs without a client-side BM25 dependency (#5851)

Bug Fixes:

Core: deleteAll() now paginates through the vector store in batches of 1000 instead of listing once, so accounts with more memories than a single page had the remainder silently left behind (#4872)
Vector Stores: Supabase list() now paginates past PostgREST’s 1000-row cap instead of stopping at the first page, search() warns when results may have been truncated by that same cap, and the initialization probe reads a row instead of writing a test vector, so Row Level Security policies that only grant read access no longer fail table verification (#6695)
Embeddings: Honor TOGETHER_API_BASE in the Together embedder, matching the Together LLM provider, so a custom gateway URL is no longer silently ignored for embeddings (#6572)

Changes:

Rerankers: RerankerFactory’s default LLM reranker model is now gpt-5-mini (was gpt-4o-mini) (#6703)

Security:

Dependencies: Patched 32 high and 57 medium severity dependency vulnerabilities across the pnpm workspace via pnpm.overrides (axios, brace-expansion, js-yaml, postcss, protobufjs, mongoose, tar, fast-xml-parser, thrift) (#6639)

v3.1.2

Bug Fixes:

Vector Stores: Apply every operator in a Cassandra compound field filter (e.g. { age: { gte: 10, lte: 20 } }) instead of stopping after the first, so the remaining bounds are no longer silently ignored (#6511)
Vector Stores: Stop the Chroma where-clause translator from dropping filter conditions. Same-field ranges (gte + lte), multi-field conditions inside $or, and negated contains/icontains under $not each collapsed to a single clause or vanished, widening the search instead of narrowing it (#6521)
Vector Stores: Skip "*" wildcard filter values in Milvus instead of matching them literally, so a filter like { user_id: "*" } no longer returns zero memories (#6508)
Vector Stores: Read textLemmatized for BM25 keyword search on Milvus, OpenSearch, and MongoDB, matching the field the memory layer actually writes, so hybrid search on those backends no longer loses the keyword signal (#6497)
LLMs: Forward responseFormat to Gemini’s responseMimeType in generateResponse(), so requesting json_object returns JSON instead of free-form text (#6468)
LLMs: Find the Anthropic text block by type instead of indexing content[0], so a thinking-enabled model whose thinking block comes first no longer throws Unexpected response type from Anthropic API (#6506)

v3.1.1

New Features:

Embeddings: Add an AWS Bedrock embedding provider (#6185)

Bug Fixes:

Packaging: Finish the lazy-loading work started in v3.1.0. The remaining LLMs (Anthropic, Google, Groq, LangChain, Mistral, Ollama), embedders (Google, LangChain, Ollama, Vertex AI), vector stores (Azure AI Search, Azure MySQL, Baidu, LangChain, Qdrant, Redis, Supabase, Valkey, Vectorize), and the Supabase history store still imported their SDKs at module load, so importing mem0ai/oss required every provider package to be installed (#6389)
Vector Stores: Convert Baidu Mochow’s raw L2 distance into a similarity score in search() (1 / (1 + distance)), so closer matches rank higher instead of lower. A row the backend returns without a score is now left undefined instead of being treated as the closest match (#6485)
Memory (OSS): Coerce non-string entity IDs (e.g. a numeric user_id) to strings instead of crashing on .trim() (#6263)
Memory (OSS): Stop update() metadata from overwriting or injecting user_id, agent_id, run_id, or actor_id (in either snake_case or camelCase). These identity fields are immutable after creation, so passing them in metadata can no longer move a memory into a different tenant’s scope (#6343)
Vector Stores: Scope Pinecone deleteCol()/reset() to the configured namespace instead of deleting the whole index, so resetting a namespaced Pinecone store no longer wipes out the other namespaces sharing that index (#6287)

Changes:

Client: Remove the unused retrievalCriteria field from PromptUpdatePayload. It was accepted and forwarded but never affected retrieval, so removing it is not a behavior change (#6313)

v3.1.0

The largest provider release for the TypeScript OSS SDK so far: 17 new vector stores, 5 new LLM providers, 4 new embedders, and reranking support. Importing mem0ai/oss no longer pulls in any provider SDK, so you only install what you actually configure.New Features:

Rerankers: Add reranking to the OSS SDK with four providers (Cohere, ZeroEntropy, cross-encoder, and LLM-based), plus per-search rerank via a rerank option on search() (#6055)
Memory (OSS): Accept text in Memory.update(). data still works but is now deprecated, so prefer text in new code (#6044)
Vector Stores: Add Pinecone (#5802), Weaviate (#5800), Milvus (#5889), Chroma (#6145), MongoDB (#5793), Elasticsearch (#5866), and OpenSearch (#5810)
Vector Stores: Add Databricks (#5824), AWS Neptune Analytics (#5797), S3 Vectors (#5822), Azure MySQL (#5827), and Google Vertex AI Vector Search (#5791)
Vector Stores: Add Turbopuffer (#5801), Upstash Vector (#5811), Valkey (#5826), Cassandra (#5823), and Baidu Mochow (#5790)
LLMs: Add AWS Bedrock (#5890), xAI Grok (#6115), Together (#6049), vLLM (#5805), and Sarvam (#6130)
Embeddings: Add Vertex AI (#5882), HuggingFace (#6027), FastEmbed (#5862), and Together (#5989)

Improvements:

Packaging: Lazy-load optional provider SDKs so importing mem0ai/oss never requires them. Provider packages are now resolved on first use, so an app that only configures OpenAI and Qdrant does not need the other provider SDKs installed (#6280)

Bug Fixes:

Memory (OSS): Re-raise LLM extraction transport failures instead of returning [], so a network error during extraction surfaces as an error rather than a silently empty result (#6102)
Vector Stores: Prevent an unhandled promise rejection in the Supabase and Redis constructors (#6111)
Client: Encode dynamic URL path segments so IDs containing special characters no longer produce malformed requests (#5963)

Security:

Dependencies: Patch the fast-xml-parser and tar transitive CVEs (#6160)

v3.0.13

Bug Fixes:

Embeddings: Guard against an embed_batch count mismatch in the OpenAI and Azure OpenAI embedders (#5966)
LLMs: Handle empty Google chat candidates without crashing (#5817)

v3.0.12

New Features:

Client: Add expirationDate to AddMemoryOptions, update(), and the Memory interface; add showExpired to SearchMemoryOptions and GetAllMemoryOptions (#5874)
LLMs: Add MiniMaxLLM provider backed by the OpenAI-compatible MiniMax API (api.minimax.io/v1, default model MiniMax-M2.7) (#5858)
LLMs: Add LiteLLM provider for routing requests through a local or hosted LiteLLM proxy (#5830)
Vector Stores: Add connectionString and ssl options to the PGVector config, allowing connection via URI instead of individual host/user/password/port fields (#5789)

Bug Fixes:

Memory (OSS): Validate and trim entity IDs (userId, agentId, runId) in deleteAll() via validateAndTrimEntityId (#5735)
Vector Stores: Use nullish coalescing for hash and timestamps in the Redis insert() and update() paths so entity payloads that omit those fields no longer crash (#5860)

Security:

Dependencies: Bump undici to >=6.27.0 via pnpm override to remediate CVE-2026-12151 (#5861)

v3.0.11

Bug Fixes:

Memory (OSS): Align entity extraction with Python by reducing generic entity noise, preserving useful topic phrases, and exact-deduplicating entity links before semantic matching (#5829)

v3.0.10

Bug Fixes:

Memory (OSS): Guard against malformed image_url entries in parseVisionMessages to prevent crashes (#5631)
Memory (OSS): Return attributedTo from get(), search(), and getAll() (#5675)
Memory (OSS): Preserve message roles in the extraction input so assistant facts aren’t attributed to the user (#5643)
Memory (OSS): Reject empty or blank messages in Memory.add() to prevent hallucinated memories (#5545)
Memory (OSS): Check message.role instead of content when detecting system messages (#3921)
LLMs: Honor the configured baseURL in AnthropicLLM (#5740)
Client: Preserve customCategories names through key conversion (#5741)
Client: Prevent hallucinated memories on an empty messages payload (#5613)
Client: Preserve user metadata keys across the case-conversion round-trip (#5515)

Security:

Dependencies: Upgrade form-data to >=4.0.6 across pnpm workspaces to remediate CVE-2026-12143 (#5618)

v3.0.9

Bug Fixes:

LLMs: Fix Anthropic tool_choice format: was incorrectly sent as a bare string "auto" (rejected by the API); now correctly sent as { type: "auto" }. Also fixes tool response parsing: tool_use blocks are now parsed into toolCalls objects instead of throwing. Updated default model to claude-sonnet-4-6 and default max_tokens to 2000 to match the Python provider. Added temperature, topP, and maxTokens to LLMConfig so Anthropic params can be configured (#5537)
Memory (OSS): Preserve custom metadata fields during update(): fields such as category, priority, and other user-defined keys were previously dropped on update; the existing payload is now spread before applying the new data (#5480)
Client: Preserve user-defined schema keys in createMemoryExport (#5594)

Security:

Dependencies: Bump esbuild to >=0.28.1 across all npm packages via pnpm overrides to remediate upstream vulnerability (#5563)

v3.0.8

New Features:

Memory: Add a contextual OSS-to-Platform notices system that surfaces occasional, situation-aware messages (first run, scale/performance thresholds, slow queries, and when temporal/decay features are relevant) pointing to the corresponding Mem0 Platform capabilities; disable via MEM0_TELEMETRY=false (#5494)

Security:

Dependencies: Upgrade @langchain/community to ^1.1.18 to remediate CVE-2026-27795 and CVE-2026-26019 (#5510)
Dependencies: Resolve all open MEDIUM Dependabot alerts via pnpm overrides (#5489)

v3.0.7

New Features:

Embeddings: Add LMStudioEmbedding provider for local embeddings via the LM Studio server (#5377)
Memory: Add opt-in explain: true option to Memory.search(). When enabled, each result includes a scoreBreakdown object with semantic, keyword, entityBoost, and temporalBoost fields so callers can inspect and tune retrieval ranking (#5102)

Bug Fixes:

Memory: Parallelize entity boost searches in Memory.search(). All entity embed + store lookups now run concurrently instead of sequentially, eliminating multi-second latency on entity-rich queries with remote embedding providers (#5377)
Vector Stores: Normalize similarity scores to [0, 1] (higher = better): fixed score inversion in the Redis vector store adapter (#5391)
Embeddings: Request encoding_format: "float" from the OpenAI embedder in both embed() and embedBatch(). Fixes incorrect vector dimensions when using OpenAI-compatible proxies that default to base64 encoding (#5170)

v3.0.6

Security:

Dependencies: Bumped axios to ^1.16.0 to remediate high-severity prototype-pollution CVEs (credential theft, MITM, DoS). Pinned transitive dependencies via pnpm overrides: jws → 4.0.1 (CVE-2025-65945), langsmith → ^0.6.0 (CVE-2026-45134), tar-fs → ^2.1.4 (CVE-2025-48387, CVE-2025-59343), picomatch → ^2.3.2 (CVE-2026-33671), minimatch → ^3.1.3 / ^5.1.8 / ^9.0.7 (CVE-2026-27903, CVE-2026-27904, CVE-2026-26996), path-to-regexp → ^8.4.0 (CVE-2026-4926), rollup → ^4.59.0 (CVE-2026-27606), glob → ^10.5.0 (CVE-2025-64756), @modelcontextprotocol/sdk → ^1.25.4 (CVE-2025-66414, CVE-2026-0621)

v3.0.5

New Features:

Client: delete() accepts an options object with deleteLinked (serialized as delete_linked, default false). When true, deleting a memory also removes the older memories it superseded (the v3 linked chain), transitively: the delete-side counterpart of latestOnly, so a superseded memory does not resurface after the current one is deleted (#5270)

v3.0.4

Bug Fixes:

Vector Stores: PGVector adapter now supports rich filter operators (eq, ne, gt, gte, lt, lte, in, nin, contains, icontains, wildcard *, $or, $not) in search(), keywordSearch(), and list(). Previously only exact-equality filters worked: operator objects were passed as raw values and returned incorrect results (#5263)

v3.0.3

Bug Fixes:

Telemetry: Stitch OSS and platform PostHog identities on MemoryClient init so $identify events fire and a single user is no longer tracked as two or three disconnected personas (#5040)
Vector Stores: Fix inverted vector distance in PGVector implementation (#4944)
Security: Harden against SQL injection and prompt injection (#4997)

New Features:

SDK: Expose decay on project.update (#5062)

v3.0.2

Bug Fixes:

LLMs: Forward timeout config to OpenAI client in JS OSS LLM providers (#4770)

Improvements:

Telemetry: Harden TS telemetry version injection and require changelog entry on version bump (#4900)
Docs: Update memory tool list, CLI usage, and config file reading logic (#4861)

v3.0.1

Bug Fixes:

Telemetry: SDK version is now injected into telemetry at build time via esbuild’s define, replacing the two hardcoded version strings in src/client/telemetry.ts and src/oss/src/utils/telemetry.ts. Previously these were stuck at 2.1.36 and 2.1.34 while the published package was on 3.x, so every telemetry event was reporting the wrong client_version. The placeholder is substituted with a string literal at bundle time: no runtime require("./package.json") in the shipped bundle (#4897).

v3.0.0

Major Release: TypeScript SDK with V3 memory pipeline, camelCase parameters, and cleaned-up API surface.V3 Memory Pipeline (OSS):

Single-Pass Extraction: Additive extraction pipeline aligned with Python SDK: memories accumulate, no UPDATE/DELETE events (#4805)
Entity Extraction & Linking: New entity_extraction.ts module (720+ lines) with cross-memory relationship retrieval (#4805)
Message Persistence: SQLite-based message history via new SQLiteManager.ts with rolling window for LLM context (#4805)
Batch Embeddings: embedBatch() support in OpenAI and Azure embedding providers (#4805)
Scoring & Lemmatization: New scoring.ts and lemmatization.ts utilities for hybrid search (#4805)
New Prompts: prompts/index.ts (592+ lines) with additive extraction prompt aligned with Python SDK (#4805)
V3 API Endpoints: MemoryClient.add() now posts to /v3/memories/add/; MemoryClient.getAll() posts to /v3/memories/ with paginated envelope { count, next, previous, results } (#4856)
Default model: gpt-5-mini is now the default in OpenAI, OpenAIStructured, and Azure LLM providers (#4829)

Breaking Changes:

External Graph Store Removed (OSS): graph_memory.ts (675 lines), graphs/tools.ts (267 lines), graphs/utils.ts (116 lines), graphs/configs.ts (30 lines) deleted. The external graph store integration is no longer part of the OSS SDK; graph memory now runs natively as built-in entity linking (#4805)
camelCase Parameters (Client SDK): All user-facing parameters converted from snake_case to camelCase. Mapping is transparent at API boundary via camelToSnakeKeys() / snakeToCamelKeys() (#4776)
// Before client.add(messages, { user_id: "alice", top_k: 5 }); // After client.add(messages, { userId: "alice", topK: 5 });
Per-Method Option Types: Replaced monolithic MemoryOptions with typed interfaces: AddMemoryOptions, SearchMemoryOptions, GetAllMemoryOptions, DeleteAllMemoryOptions (#4740)
Removed Deprecated Parameters: org_id, project_id, api_version, output_format, async_mode, enable_graph, limit removed from client method signatures. ClientOptions reduced to { apiKey, host } only (#4740)
limit renamed to topK (OSS): Update all search calls (#4740)
topK default changed 100 → 20 in Memory.getAll() and Memory.search(). Pass topK: 100 explicitly to restore the old behavior (#4843)
Entity ID validation: userId / agentId / runId are trimmed; empty-string and whitespace-only values now throw (#4843)
Search params validation: threshold must be in [0, 1]; topK must be a non-negative integer: invalid inputs throw (#4843)
messages in Memory.add() is required: Passing undefined or null now throws (#4843)
customPrompt renamed to customInstructions (OSS): Update memory and vector store configurations (#4740)
enableGraph removed (OSS): Config option removed: graph memory no longer available in OSS (#4776)

New Features:

LLMs: Added DeepSeek LLM provider with OpenAI-compatible integration using custom baseURL to api.deepseek.com (#4613)
Entity store isolation: MemoryVectorStore now uses a dedicated _entities.db file, preventing entity/memory store collisions (#4829, #4841)
Payload backward compatibility: Legacy camelCase payload keys normalized to snake_case on read (#4841)

Bug Fixes:

V3 migration: Fixed crashes in the OSS migration path; entity linking works end-to-end (#4836)
PGVector init race: PGVector.initialize() now memoises the in-flight init promise (#4841)
Redis module detection: Handles both node-redis v4+ and legacy moduleList response shapes (#4841)
Config: Fixed ConfigManager.mergeConfig() to only include graphStore when explicitly provided by user, preventing default Neo4j connection attempts (#4776)
LLMs: Config manager now falls back to userConf.url for baseURL: prevents custom LLM providers (Ollama, LMStudio) from silently connecting to OpenAI (#4761)

Improvements:

Telemetry: Sample OSS hot-path events at 10% to reduce PostHog event volume (#4771)

See the OSS v2 to v3 migration guide for upgrade instructions.

v2.4.6

New Features & Updates:

Client: Added multilingual parameter to project update types (#4314)

v2.4.5

Bug Fixes:

OSS: Replace .single() with .maybeSingle() in SupabaseDB.get() to handle missing rows (#4599)
Embeddings: Pass dimensions parameter to OpenAI embeddings API (#4632)
OSS: Extract JSON from chatty LLM responses in fact retrieval (#4533)

v2.4.4

Bug Fixes:

OSS: Fixed Qdrant Cloud “Illegal host” error by defaulting to port 6333 when URL has no explicit port (#4565)

v2.4.3

New Features & Updates:

OSS: Added pgvector support to NodeJS OSS VectorStoreFactory (#3997)

Bug Fixes:

OSS: Made pgvector pg import compatible with ESM (#4544)
OSS: Registered pgvector in VectorStoreFactory (#4502)
OSS: Used root LLM config as fallback for graph store instead of hardcoded OpenAI default (#4466)
OSS: Fixed toCamelCase in Redis get method for the payload (#3172)
Client: Fixed Zod Schema incompatibility with OpenAI Structured Outputs API (#3462)

v2.4.2

Bug Fixes:

Client: Fixed webhook createWebhook and updateWebhook API serialization
Client: Added missing MEMORY_CATEGORIZED event type to WebhookEvent enum
Types: Added WebhookCreatePayload and WebhookUpdatePayload for better type safety

Tests:

Added end-to-end unit test coverage for the platform client: CRUD, batch, search, webhooks, users, project, and initialization (#4357)
Added real API integration tests for memory CRUD, batch operations, search, user management, project configuration, and webhook lifecycle (#4395)
Deleted obsolete e2e test files replaced by the new structured test suite (#4419)

v2.4.1

Bug Fixes:

Core: Fixed code block content extraction: content inside code blocks is now properly extracted instead of being deleted (#4317)

Improvements:

Code Quality: Fixed linting issues across the SDK (#4334)

v2.4.0

Bug Fixes:

OSS Storage: Fixed SQLITE_CANTOPEN errors when running as a LaunchAgent, systemd service, or in containers where process.cwd() is read-only (e.g. /). Default vector_store.db location changed from process.cwd()/vector_store.db to ~/.mem0/vector_store.db.
OSS Storage: Fixed historyDbPath config being silently ignored: config merging always overwrote it with defaults. Top-level historyDbPath is now correctly propagated into historyStore.config with proper precedence.
OSS Storage: Added ensureSQLiteDirectory(): parent directories for SQLite database files are now auto-created before opening, preventing SQLITE_CANTOPEN when using nested paths.

Improvements:

Migration: Added deprecation warning when an existing vector_store.db is found at the old process.cwd() location, guiding users to move it or set vectorStore.config.dbPath explicitly.
Config: Limited default SQLite config spreading to only SQLite history providers, preventing config leaking into Supabase or other providers.

v2.3.0

Breaking Changes:

Dependencies: Minimum Node.js version for OSS sqlite features is now Node 20+ (due to better-sqlite3 v12)

Bug Fixes:

OSS Storage: Replaced sqlite3 with better-sqlite3 to fix native binding resolution failures under jiti-based loaders (e.g. OpenClaw plugin system). Fixes issues where the bindings module walked V8 stack frames with synthetic filenames, failing to locate the native .node addon.
OSS Storage: Fixed async init race condition in SQLiteManager: init() is now synchronous
OSS Vector Store: Migrated MemoryVectorStore from sqlite3 to better-sqlite3 with transactional batch inserts

Improvements:

Performance: Cached prepared statements in SQLiteManager for faster history operations
Performance: Batch insert() in MemoryVectorStore wrapped in a transaction for atomicity
Build: Updated tsup.config.ts externals from sqlite3 to better-sqlite3

v2.2.3

New Features & Updates:

Memory Update:
- Added timestamp parameter to update(): accepts Unix epoch or ISO 8601 string

v2.2.2

New Features & Updates:

Project Settings:
- Added inclusion prompt, exclusion prompt, memory depth, and usecase setting

v2.2.1

Improvements:

Client: Added support for keyword arguments in add and search methods, allowing additional properties beyond defined options for experimental features

v2.2.0

New Features:

Vector Stores: Added Azure AI Search vector store support

Improvements:

Config: Fixed embedder config schema to support embeddingDims and url parameters
Graph Memory: Replaced hardcoded LLM provider with provider from configuration

Bug Fixes:

Embedders: Fixed hardcoded embeddingDims values in embedders (OpenAI, Ollama, Google, Azure)
Build: Fixed TypeScript build errors

v2.1.38

New Features:

Client: Added metadata param to update method.

v2.1.37

New Features:

OSS: Added RedisCloud search module check

v2.1.36

New Features:

Client: Added structured_data_schema param to add method.

v2.1.35

New Features:

Client: Added createMemoryExport and getMemoryExport methods.

v2.1.34

New Features:

OSS: Added Gemini support

v2.1.33

Improvement:

Client: Added immutable param to add method.

v2.1.32

Improvement:

Client: Made api_version V2 as default.

v2.1.31

Improvement:

Client: Added param filter_memories.

v2.1.30

New Features:

OSS: Added Cloudflare support

Improvements:

OSS: Fixed baseURL param in LLM Config.

v2.1.29

Improvements:

Client: Added Async Mode Param for add method.

v2.1.28

Improvements:

SDK: Update Google SDK Peer Dependency Version.

v2.1.27

Improvements:

OSS: Added baseURL param in LLM Config.

v2.1.26

Improvements:

Client: Removed type string from messages interface

v2.1.25

Improvements:

Client: Improved error handling in client.

v2.1.24

New Features:

Client: Added new param output_format to match Python SDK.
Client: Added new enum OutputFormat for v1.0 and v1.1

v2.1.23

New Features:

Client: Updated deleteUsers to use v2 API.
Client: Deprecated deleteUser and added deprecation warning.

v2.1.22

New Features:

Client: Updated deleteUser to use entity_id and entity_type

v2.1.21

Improvements:

OSS SDK: Bumped version of @anthropic-ai/sdk to 0.40.1

v2.1.20

Improvements:

Client: Fixed organizationId and projectId being assigned to default in ping method

v2.1.19

Improvements:

Client: Added support for timestamps

v2.1.18

Improvements:

Client: Added support for custom instructions

v2.1.17

New Features:

OSS SDK: Added support for Langchain LLM
OSS SDK: Added support for Langchain Embedder
OSS SDK: Added support for Langchain Vector Store
OSS SDK: Added support for Azure OpenAI Embedder

Improvements:

OSS SDK: Changed model in LLM and Embedder to use type any from string to use langchain llm models
OSS SDK: Added client to vector store config for langchain vector store
OSS SDK: - Updated Azure OpenAI to use new OpenAI SDK

v2.1.16-patch.1

Bug Fixes:

Azure OpenAI: Fixed issues with Azure OpenAI

v2.1.16

New Features:

Azure OpenAI: Added support for Azure OpenAI
Mistral LLM: Added Mistral LLM integration in OSS

Improvements:

Zod: Updated Zod to 3.24.1 to avoid conflicts with other packages

v2.1.15

Improvements:

Client: Added support for Mem0 to work with Chrome Extensions

v2.1.14

New Features:

Mastra Example: Added Mastra example
Integrations: Added Flowise integration documentation for Mem0 memory setup

Improvements:

Demo: Updated Demo Mem0AI
Client: Enhanced Ping method in Mem0 Client
AI SDK: Updated AI SDK implementation

v2.1.13

Improvements:

Introduced ping method to check if API key is valid and populate org/project id

AI SDK v1.0.0

New Features:

Vercel AI SDK Update: Support threshold and rerank

Improvements:

Made add calls async to avoid blocking
Bump mem0ai to use 2.1.12

v2.1.12

New Features:

Mem0 OSS: Support infer param

Improvements:

Updated Supabase TS Docs
Made package size smaller

v2.1.11

New Features:

Supabase Vector Store Integration
Feedback Method

Python v0.2.13 / Node v0.2.14

Improvements:

Client: Requests now carry the three surface-identity headers (X-Mem0-Source, X-Application, X-Mem0-Client) introduced in the Python and TypeScript SDKs, so the platform can attribute calls made through the CLI to the correct surface and version (#7326)

Python v0.2.12 / Node v0.2.13

New Features:

version: New mem0 version subcommand, alongside the existing --version flag, so scripts and agent harnesses can read the CLI version as a regular subcommand instead of a root-level flag (Python and Node #6907)
add: New --agent-custom-instructions flag, threaded through to agent_custom_instructions on the /v3/memories/add/ payload: a second extraction instruction set that applies only to agent-scoped memories, matching the SDKs’ agentCustomInstructions / agent_custom_instructions support (Python and Node #6910)

Documentation:

search --filter: The --filter help text and docs/platform/cli.mdx now spell out the JSON shape ({"AND": [...]} / {"OR": [...]}) with a concrete example instead of just calling it “an advanced filter expression,” and a matching example command was added to both the CLI help text and the docs page (Python and Node #6907)

Python v0.2.11 / Node v0.2.12

New Features:

add: New --custom-instructions, --custom-categories, --structured-data-schema, and --timestamp flags, matching the Platform /v3/memories/add/ payload fields (Python and Node #6696)
search: New --show-expired, --reference-date, and --latest-only flags, forwarded to /v3/memories/search/ as show_expired, reference_date, and latest_only (Python and Node #6696)
list: New --show-expired and --latest-only flags (Python and Node #6696)
update: New --expires and --timestamp flags (Python and Node #6696)
delete: New --delete-linked flag to also delete memories linked to the target memory (Python and Node #6696)

Bug Fixes:

help --json: Emit JSON automatically under agent mode (--agent), not only when --json is passed explicitly. Python checks is_agent_mode() and now renders through console.print_json() instead of a plain console.print(json.dumps(...)) call; Node also checks the root --agent flag in addition to the subcommand’s --json flag (Python and Node #6773)

Changes:

add: --categories is no longer forwarded on add. The /v3/memories/add/ payload has no categories field (only custom_categories), so the CLI was sending a key the endpoint does not accept. Passing the flag now exits 1 with a message pointing to --custom-categories (Python and Node #6696)

Security:

Dependencies (Node): Tighten the postcss pnpm override from <8.5.10 → >=8.5.10 to <8.5.18 → >=8.5.18 <9.0.0, closing a newer CVE range the previous floor didn’t cover, as part of a wider dependency patch sweep across the pnpm workspaces (#6639)

Python v0.2.10 / Node v0.2.11

Bug Fixes:

Platform backend: Encode dynamic URL path segments so memory and entity IDs containing special characters no longer produce malformed requests (Python and Node #5963)

Python v0.2.9 / Node v0.2.10

Bug Fixes:

entity delete: Keep every deleted entity in the result output instead of only the last one (Python #5936, Node #5970)
Output formatters (Python): Handle null memory fields in the output formatters instead of erroring (#5957)
Config (Python): Reject invalid integer config values with a clean error instead of a traceback (#5956)

Python v0.2.8 / Node v0.2.9

Security:

Telemetry key handling: CLI telemetry no longer passes the Mem0 API key to a child process through command-line arguments. The context is sent over stdin, so the key is not visible in ps, /proc/<pid>/cmdline, or Activity Monitor. Fixes #4862.

Bug Fixes:

Python package version: __version__ now matches the packaged mem0-cli version.

Node v0.2.8

Security:

Dependencies: Pinned transitive dependencies via pnpm overrides to remediate high-severity CVEs: jws → 4.0.1 (CVE-2025-65945), langsmith → ^0.6.0 (CVE-2026-45134), tar-fs → ^2.1.4 (CVE-2025-48387, CVE-2025-59343), picomatch → ^2.3.2 (CVE-2026-33671), minimatch → ^3.1.3 / ^5.1.8 / ^9.0.7 (CVE-2026-27903, CVE-2026-27904, CVE-2026-26996), path-to-regexp → ^8.4.0 (CVE-2026-4926), rollup → ^4.59.0 (CVE-2026-27606), glob → ^10.5.0 (CVE-2025-64756), @modelcontextprotocol/sdk → ^1.25.4 (CVE-2025-66414, CVE-2026-0621)

Python v0.2.7 / Node v0.2.7

New Features:

mem0 whoami: Prints the active agent’s default_user_id from local config without a network call.
mem0 agent-rush <add | search>: Wraps the /v1/agent-rush/ platform endpoints for the AGENTRUSH game. Project routing is resolved server-side, and platform error codes are shown as actionable hints.
PII safety prompt: The first interactive mem0 agent-rush add asks users to acknowledge that AGENTRUSH memories are public. The acknowledgement is stored in ~/.mem0/config.json under agent_rush.acknowledged_at.
Config schema: Added agent_rush.acknowledged_at for the local acknowledgement timestamp.

Changes:

Agent Rush telemetry: Agent Rush commands send X-Mem0-Mode: agent-rush so platform telemetry can split game traffic from regular CLI usage.

Python v0.2.6 / Node v0.2.6

Bug Fixes:

Claim flow error message: The email_already_claimed tip in mem0 init --email previously suggested running mem0 link <key>: a command that doesn’t exist. Replaced with honest copy pointing the user to sign in at app.mem0.ai with their existing credentials (#5152)

Python v0.2.5 / Node v0.2.5

New Features:

Agent Mode (mem0 init --agent): Zero-friction signup for AI agents: mints a working Mem0 API key in under 5 seconds with no email, no dashboard, no OTP. Returns an unclaimed shadow account the human can later claim with mem0 init --email <their-email> (memories preserved, same key keeps working) (#5123)
Self-declared agent identity: Agents pass --agent-caller <name> (e.g. claude-code, cursor, codex) on mem0 init --agent so signups attribute to the right tool in analytics. Proof Editor-style: the agent declares itself rather than the CLI sniffing it from env vars (#5123)
mem0 identify <name>: New subcommand to self-tag an Agent Mode key after the fact when the agent forgot to pass --agent-caller on init. Idempotent: re-running just overwrites (#5123)
Plugin sync: ~/.claude/settings.json::env::MEM0_API_KEY and ~/.zshrc/.bashrc export MEM0_API_KEY= lines stay in sync with ~/.mem0/config.json automatically. Idempotent: only updates EXISTING entries, never creates new ones (#5123)
Claim flow: mem0 init --email <email> claims an existing Agent Mode shadow via OTP. Upgrade-in-place: the API key never changes, memories transfer to the human’s account (#5123)

Bug Fixes:

Decision tree network resilience: pingKey now distinguishes network errors from invalid keys: returns false ONLY on HTTP 401/403, returns true on connection failures / timeouts / 5xx. Prevents a VPN flap from silently rotating the user’s API key and rewriting plugin-sync targets (#5123)
Rate-limit error clarity: DRF’s opaque "You do not have permission" 403 from Agent Mode rate limits is now translated to "Daily Agent Mode signup limit reached for this network (5/day). Try again from a different IP or after midnight UTC." (#5123)
JSON envelope command field: mem0 init --agent --json error envelopes now populate the command field correctly instead of returning an empty string (#5123)
Bootstrap envelope validation: Defends against partial/malformed backend responses (e.g. {api_key: null}) silently persisting null/undefined into typed string fields (#5123)

Python v0.2.4 / Node v0.2.4

New Features:

V3 API Routes: Migrated add, search, and list commands from v1/v2 to v3 API endpoints: POST /v3/memories/add/, POST /v3/memories/search/, POST /v3/memories/. Aligns both CLIs with the Python and TypeScript SDKs which already use v3 (#4916)

Breaking Changes:

--graph / --no-graph removed: The enable_graph config option, --graph and --no-graph CLI flags, and MEM0_ENABLE_GRAPH environment variable have been removed from both CLIs. Graph memory is now a project-level setting on the Platform (#4916)

Python v0.2.3 / Node v0.2.3

Bug Fixes:

Telemetry: Replaced shared "anonymous-cli" fallback with a persistent per-machine random hash (cli-anon-<uuid>), so anonymous CLI users are counted individually in PostHog instead of collapsing into one identity (#4789)
Telemetry: Added PostHog $identify event on first authenticated run to stitch pre-signup anonymous history onto the authenticated user profile (#4789)

Improvements:

API: All API calls now include source=CLI in request bodies (POST/PUT) and query params (GET/DELETE) for server-side attribution (#4789)

Python v0.2.2 / Node v0.2.2

New Features:

Telemetry: Added PostHog telemetry and source tracking to both Python and Node CLIs (#4699)
Validation: API key validated upfront via /v1/ping/ on startup: fail-fast with a helpful error instead of cryptic 401s (#4701)

Bug Fixes:

CD: Fixed OIDC trusted publishing with npx npm@latest (#4724)
CD: Removed npm self-upgrade from CD workflows (#4723)

Python v0.2.1 / Node v0.2.1

New Features:

Docs: Comprehensive README with installation, usage examples, and purple branding (#4680)

Bug Fixes:

npm: Added repository field to Node packages for npm provenance (#4671)
CD: Added CD workflows for Node SDK packages with OIDC trusted publishing (#4670)

Python v0.2.0 / Node v0.1.1

New Features:

event commands: mem0 event list shows recent background processing events in a table; mem0 event status <id> shows full detail including nested memory results (#4649)
--json / --agent flag: Root-level flag switches all command output to a structured JSON envelope for programmatic/agent consumption. Envelope format: {"status", "command", "duration_ms", "scope", "count", "data"} (#4649)
Agent output sanitization: Raw API responses projected to only relevant fields per command (e.g., add → {id, memory, event}, search → {id, memory, score, created_at, categories}) (#4649)
Email login: Added email verification code login to mem0 init (#4623)
Brand update: Updated color palette from purple to golden (#4664)
CI/CD: Added CI pipelines and CD workflows for both CLIs (#4640, #4653)

Bug Fixes:

Node: Fixed critical MODULE_NOT_FOUND crash on status, import, and all commands when installed globally: replaced runtime createRequire with build-time version injection (#4636)
Node: API errors now show full response detail instead of bare “Bad Request” (#4636)
Python: Fixed double error printing on all commands (#4636)
status command: Replaced heavyweight /v1/entities/ check with dedicated GET /v1/ping/ endpoint (#4649)
add command: Deduplicated PENDING results from API; changed misleading count message (#4649)
init command: Partial flags now work in non-TTY; warns before overwriting existing config; added --force flag (#4649)
delete command: Fixed entity delete via v2 API for all entity types (#4649)

Improvements:

Tables now show full UUIDs (was truncated to 8 chars, making mem0 get <id> fail) (#4636)
Search table includes Score column (#4636)
config get api_key short-form aliases added (#4636)
Client-side validation for --expires, --page-size, --page, --top-k, --threshold, and empty content (#4636)
printInfo / printScope moved to stderr to avoid contaminating JSON piping (#4636)

Python v0.1.0 / Node v0.1.0

Initial Release: Official Mem0 CLIA full-featured command-line interface for Mem0, available in both Python and Node.js:

Install: pip install mem0-cli (Python) or npm install -g @mem0/cli (Node.js)
Full command suite: add, search, list, get, update, delete, import, config, init, status, entity
Interactive setup: mem0 init with API key entry and user ID configuration
Works everywhere: Platform (Mem0 Cloud) and self-hosted OSS modes
Scriptable: -o json flag for CI/CD pipelines and automation
Dual SDK: Same commands, same experience across Python and Node.js
Shared spec: Both implementations driven by a single cli-spec.json ensuring identical behavior (#4575)

Shared agent plugin runtime

Changed:

Consolidated the coding-agent integrations into integrations/agent-plugin-core/: one Python runtime, one TypeScript utility library, and six canonical Python-plugin skill templates. Native adapters retain each host’s event contracts and capabilities.
Python plugins ship generated, self-contained core/ and skills/ directories. Builds validate portable schemas and skills, parse native JSON, and reject generated-file drift, missing files, stale generated files, and symlinks. TypeScript packages bundle the shared source into their distributable JavaScript and verify their entry points.
Replaced the old integrations/mem0-plugin/ layout with native host directories and one portable integrations/mem0-agent-plugin/ package. Updated marketplace paths, installation guides, and integration-skill links. OpenCode now lives in integrations/opencode-plugin/.
Native Python plugins expose one local, read-only search_memories MCP tool and six skills: search, remember, forget, status, pause, and resume. The shared search tool accepts optional run_id with every scope (repo, dir, and mine) to recall memories from a specific coding-agent session. Omitting it searches across sessions. This local tool is separate from the hosted Mem0 MCP server’s tool set.

Fixes:

Hooks, controls, MCP servers, and detached workers use the same host-specific data directory. Detached workers retain the host identity and telemetry source; --plugin-data-dir reaches the shared resolver.
Session-end workers flush the conversation already captured by hooks. Repeated and concurrent response hooks no longer duplicate an answer, while identical answers after separate prompts are preserved.
Shared prompts and responses are redacted without the previous 6,000-character cutoff. Python extraction splits oversized messages without dropping text to enforce each request’s input budget. Flush event selection and claims share one write transaction, delayed handoffs are replaced atomically, and permanent HTTP polling errors fail promptly.
Extraction instructions refer to the current coding agent. Python redaction covers JSON-shaped credentials; both telemetry runtimes recursively remove sensitive keys, including keys inside nested lists.
New Git repository writes use a hash of the remote identity in agent_id. Search and explicit shared-memory deletion include both current and legacy repository IDs within the repository’s app_id. Existing memories are not rewritten. Legacy IDs retain their original ambiguity for matching owner/repository names on different Git hosts.

Packaging:

Claude Code, Cursor, Codex, Kimi, Antigravity, and the portable Python bundle are versioned at 0.3.1. OpenCode and DeepSeek Harness are 0.3.0; Pi Agent is 0.3.0; OpenClaw is 1.1.0. Each host’s changes and upgrade considerations are listed in its tab.
Python and TypeScript CI run their respective runtime suites. Package checks build the installable artifacts, check generated-file consistency, and reject TypeScript output that still imports monorepo source.

#7203

mem0-plugin v0.2.15

Fixes:

Search: A failed search request now prints [mem0] search request failed: <error> to stderr before returning no results. search_memories() swallowed every exception and returned [], so an expired API key, a network failure, or a 500 from the backend was indistinguishable from a genuine “nothing stored yet” and the agent carried on with no context and no warning. Shared by Claude Code, Cursor, Codex, Antigravity, and Kimi (#6898)
Cursor: mcpServers in .cursor-plugin/plugin.json now points at ./.cursor-mcp.json instead of .cursor-mcp.json. The un-prefixed path resolved inconsistently depending on Cursor’s working directory when it loaded the plugin (#6948)
Codex: install_codex_hooks.py now prints all six registered events (PreToolUse, SessionStart, UserPromptSubmit, PostToolUse, Stop, PreCompact) after installing, instead of a stale four-event list left over from an earlier version of the installer. The README’s hook table is corrected to match, documenting the three PreToolUse handlers and two PostToolUse handlers that were previously undocumented (#6948)

Documentation:

New Claude.ai integration page (#6948)

The Claude Code, Cursor, and Codex per-editor manifests (.claude-plugin/plugin.json, .cursor-plugin/plugin.json, .codex-plugin/plugin.json) had drifted to 0.2.13 while the .claude-plugin/marketplace.json and .cursor-plugin/marketplace.json listings had already moved to 0.2.14, so installs were pinned one release behind what the marketplace advertised. This release realigns every manifest and marketplace listing to 0.2.15.

mem0-plugin v0.2.14

Fixes:

Settings loading no longer crashes on a malformed settings.json: If ~/.mem0/settings.json parsed as valid JSON but wasn’t an object (a list or a bare string, for example), load_settings() called .items() on it and raised. resolve_config() guards only against ImportError, so the failure escaped into every hook that resolves identity, not just setup. Settings loading now keeps the defaults instead. Setup also stopped claiming it “Created” the settings file when one already existed; it only prints that when it actually wrote one. Shared across Claude Code, Cursor, Codex, and Antigravity.

New Features:

Unknown settings key warning: Setup now flags keys in settings.json that the plugin doesn’t recognize, so a typo’d setting no longer fails silently.

Removed:

Dropped the mem0_doc_search “openmemory” section now that OpenMemory has been removed from the docs site, so the skill no longer links to pages that don’t exist.

mem0-plugin v0.2.13

Fixes:

Assistant messages no longer stored as your own: The session-summary hook (fires at the end of every assistant turn) and the post-compaction hook were sending the assistant’s own message to Mem0 tagged role: "user". Because Mem0 extracts facts about the user from each message and uses role to decide who spoke, the assistant’s first-person prose was being saved as the human’s stated preferences: “I recommend we drop Redis” became User prefers dropping Redis entirely. Both hooks now send role: "assistant", so the same session is stored as Assistant recommended.... Affects Claude Code, Cursor, Codex, and Antigravity, which share these hooks.

Existing memories written by the previous versions are not rewritten. If your memories contain preferences you never expressed, delete them; the plugin will not recreate them.

mem0-plugin v0.2.12

New Features:

Query-driven prefetch: The UserPromptSubmit hook now searches Mem0 with the current prompt and injects the top matches, so relevant memories are guaranteed in context instead of relying on the agent to search for them. Opt out with MEM0_PREFETCH=false.

Improvements:

Multi-hop guidance: The decision rubric now tells the agent to run follow-up searches for multi-part or comparative questions instead of stopping after one. Tool descriptions for every memory tool were sharpened with clearer “what + when to use” guidance.

mem0-plugin v0.2.11

Fixes:

Session summary metadata: files_touched now stores a real list instead of a double-encoded JSON string, so returned memories render readable file paths across Claude Code, Cursor, Codex, and Antigravity.

Improvements:

Reranked auto-injected context: File-context, bash-error lookup, and session-resume prefetch now rerank by default. Disable with MEM0_RERANK=0, false, no, or off.

mem0-plugin v0.2.10

Fixes:

Editor telemetry attribution: Antigravity is no longer counted as Claude Code, Codex no longer falls back to the generic plugin bucket, and Cursor wrappers now pin their platform before delegating to shared hook scripts.
Per-editor plugin version: Telemetry reads the manifest for the detected platform instead of always reading the Claude plugin manifest.

New Features:

MEM0_PLATFORM override: Explicit platform markers now win over env-var auto-detection for reliable telemetry labels.

mem0-plugin v0.2.9

New Features:

File-context injection: Before supported editors read a file, Mem0 searches for memories that reference that path and injects a compact prior-work timeline.
Stop hook session summaries: Session-end hooks store structured summaries with request, decisions, learnings, next steps, and files touched.
SessionStart activity timeline: Sessions with existing memories show recent decisions, bug fixes, and session summaries on startup.

Changes:

Added shared file_context.py, capture_session_summary.py, and session_timeline.py helpers for the hook pipeline.

mem0-plugin v0.2.8

New Features:

Global search mode: global_search lets reads search all users and projects in the platform project while writes stay scoped to the current user_id and app_id.
Automatic coding categories: The coding taxonomy is installed in the background on session start and gated by an API-key and taxonomy fingerprint.
Global scope controls: /mem0:switch-project --global and /mem0:switch-project --no-global persist global-search behavior in ~/.mem0/settings.json.
Scope-aware startup: Session banners show scope=global when global search is active, and startup memory counts use the same global filter.
Category setup state: ~/.mem0/categories_setup.json caches setup per Mem0 account and taxonomy fingerprint, so categories are applied once and re-applied only when the taxonomy changes.
Tests: Added coverage for category fingerprints, state-file gating, state save/load, and idempotent category application.

Changes:

Session start spawns auto_setup_categories.py in the background alongside project-file auto-import.
Onboarding no longer asks users to install coding categories manually; it verifies status and applies categories only if the background run has not finished.
New-project startup copy now tells users that coding categories install automatically.

mem0-plugin v0.2.7

Fixes:

First-install auth: Removed authorizationUrl from .mcp.json so Claude Code uses the MEM0_API_KEY header instead of falling into a broken OAuth reconnect flow.
Onboarding auth copy: /mem0:onboard no longer tells users to complete browser OAuth.
Invalid configure command: Removed claude plugin configure mem0 references and clarified that Claude Code stores the plugin API key through the enable UI prompt.

mem0-plugin v0.2.6

Fixes:

Memory count and stats: Removed the run_id: "*" filter from read paths so memories written without run_id are visible again.
Write scoping: add_memory moved session tracking from top-level run_id to metadata.session_id, keeping new memories in the default query partition.
Session stats: Memory add/search tracking moved into enforce_metadata_defaults.sh, because plugin MCP tools do not trigger PostToolUse hooks.
Auto-capture reliability: Added direct API auto-capture every third message plus a proportional prompt nudge when few memories have been stored.
Desktop API keys: _identity.sh and _identity.py can read MEM0_API_KEY from common shell profiles when Claude Code Desktop does not inherit shell env vars.
Import cleanup: Stale auto-import chunks now delete through the v1 delete endpoint, matching Mem0 v3 add behavior.

Removed:

Removed Stop and SessionEnd hook capture paths, redundant post-commit/task/subagent scripts, dead settings, and the deleted file-read test.

Changes:

Stats and tour now query with user_id + app_id filters only.
Session banners now tell agents to store learnings incrementally instead of waiting until session end.
Session-start resets per-user message counters so nudges begin cleanly.

mem0-plugin v0.2.5

Fixes:

Hook schema: PostToolUse scripts now read tool_response, Stop hook output is schema-compliant, and SessionStart uses visible raw stdout.
Tool coverage: Memory write blocking includes MultiEdit, and metadata enforcement covers add, search, get, update, and delete-all variants.
Identity injection: add_memory, search_memories, get_memories, and delete_all_memories receive safe default user_id and app_id values when the agent omits them.
Skill API calls: Export, stats, pin, health, memory-reviewer, dream, tour, onboard, context-loader, and weekly stats examples now use the correct filter and top_k shapes.
Compact summary capture: capture_compact_summary.py is spawned correctly after compaction and guarded against unset identity variables.

New Features:

Stop hook analyzer: Added stop_hook_check.py to detect meaningful work before prompting memory capture.
Skill checklists: Added progress checklists to /mem0:dream and /mem0:onboard.

Changes:

enforce_metadata_defaults.sh was rewritten around handler-based dispatch and never overrides explicitly provided identity.

mem0-plugin v0.2.4

Fixes:

PostToolUse matcher: Supports both mcp__mem0__ and mcp__plugin_mem0_mem0__ names so session stats record memory operations again.
File-read noise: File-read recall now prefers metadata.files plus score threshold before falling back to basename search.
Pinning: /mem0:pin no longer uses removed immutable=True; it relies on pinned metadata/text conventions.
Project listing: /mem0:list-projects merges null-scoped and app-scoped memories so legacy records are counted.
ID lookup: /mem0:peek routes memory-id references to get_memory instead of semantic search.

New Features:

Session ID capture: SessionStart persists the Claude session ID for session-scoped memory operations.
Search controls: _search.py gained min_score, metadata filters, rerank, and threshold support.
Reranked tour and peek: /mem0:tour and /mem0:peek use reranking for better precision.

Removed:

Removed the nonexistent PostToolUseFailure hook and the superseded /mem0:protocol skill.

mem0-plugin v0.2.3

New Features:

Read hook: Added PreToolUse(Read) file recall with skip rules for non-code files, lockfiles, and node_modules.
Shared search module: Added _search.py for v3 search calls and consistent context formatting.
Settings file: Added ~/.mem0/settings.json with search, retention, confidence, debug, and capture settings.
Memory skills: Added /mem0:forget, /mem0:peek, /mem0:memory-reviewer, and /mem0:context-loader.
Compact output style: Added compact [type] content [mem0:id] memory display.
Claude userConfig: Added sensitive api_key plugin config support.

Changes:

Skills were renamed to shorter command names, bash-error hooks now run real searches, resume prompts prefetch context, and remember-intent phrases route to /mem0:remember.
API key resolution now checks CLAUDE_PLUGIN_OPTION_API_KEY, and hook timeouts increased for prompt and bash-output searches.

Removed:

Removed /mem0:dream --forget, /mem0:dream --schedule, /mem0:health --fix, and CLI config fallback.

mem0-plugin v0.2.2

Fixes:

Compaction writes: Prevented duplicate pre-compaction captures and shortened session-state content before letting Mem0 infer structured memories.
Error detection: Reduced false positives from routine test and linter output.
Git commit noise: Removed automatic background commit-context writes and gated post-commit prompts behind settings.
Project identity: Added git-remote hash fallback when folders are renamed.
Dependency install race: Added a lock and failure sentinel for ensure_deps.sh.
Telemetry: Switched distinct_id from MD5 to SHA-256, protected system props, and loaded plugin version from the manifest.

New Features:

Chunking and import: Added shared chunking helpers and Markdown chunk imports.
Metadata defaults: Added enforce_metadata_defaults.sh for default confidence, source, and type on add_memory.
Project utilities: Added list-projects, all-project tour, weekly stats, dream forget mode, and competing-tools import.

Removed:

Removed on_pre_commit.py and merged digest, forget, import-tools, and peek flows into existing skills.

mem0-plugin v0.2.1

New Features:

Skill suite: Added 11 skills for dream, export, import, tool import, forget, health, peek, pin, remember, stats, and digest.
Hook suite: Added bash-error recall, git/commit context hooks, compaction recovery, session-end capture, subagent reminders, tool-failure recovery, and pre-commit capture.
Telemetry: Added sampled PostHog telemetry with no memory content or API keys and MEM0_TELEMETRY=false opt-out.
Coding categories: Added 10 more coding categories and dependency-managed SDK install through ensure_deps.sh.
Project onboarding: Added auto-onboarding, mem0.md config loading, inactive API-key banner, import/export parsers, and competing-tool import.

Changes:

Memory scoping moved from metadata.project_id to top-level app_id, and write paths moved to /v3/memories/add/.
Session stats gained category counters and recent memory IDs.
Tour, onboard, prompt hooks, and MCP guidance were updated for project-aware workflows.

Fixes:

Shared scripts now resolve plugin API-key env vars consistently and pass hook CWD into project/branch resolution.

mem0-plugin v0.2.0

New Features:

Project-scoped memories: Deterministic project IDs from git remotes isolate memories per repo.
Branch-aware tagging: Session-state and compact-summary memories include branch metadata.
Project file import: SessionStart imports CLAUDE.md, AGENTS.md, .cursorrules, .windsurfrules, and mem0.md as project-profile memories.
Active identity banner: Startup shows user, project, branch, and memory count.
Session report: Stop hook reports memory writes, retrieves, and touched categories.
Onboarding, tour, and switch-project skills: Added first-run setup, proof-of-value browsing, and manual project override.
Session stats tracker: Added local add/search tracking for reports.

mem0-plugin v0.1.3

Fixes:

User ID regression: user_id resolution no longer derives from MEM0_API_KEY, restoring MEM0_USER_ID to $USER to default order so older memories remain reachable.

Upgrade Notes:

The ~/.mem0/identity.json cache is no longer used.
Users who stored memories under the short-lived API-key hash can recover them by temporarily setting MEM0_USER_ID to the old hash.

mem0-plugin v0.1.2

New Features:

Identity resolver: Added deterministic user-id resolver, later reverted in v0.1.3.
Compact summaries: Added SessionStart compact handler for compact_summary memories.
Coding taxonomy setup: Added one-shot coding category setup for architecture, anti-patterns, task learnings, tooling, bug fixes, conventions, and preferences.
Debug logging: Added opt-in hook logs with MEM0_DEBUG=1.
Shared MCP skill: Replaced the Claude-Code-specific Codex skill with mem0-mcp.

Fixes:

Session IDs are stored in memory metadata, startup exits quietly without an API key, memory-write blocking avoids docs/memory/*.md, duplicate compact writes were removed, and hook captures now expire after 90 days.

mem0-plugin v0.1.1

Fixes:

Cursor plugin became fully functional.
Codex plugin support and integration docs were added.
Codex lifecycle hooks became available through the opt-in installer.
Invalid Claude plugin config keys were removed.

mem0-plugin v0.1.0

Mem0 Plugin for Claude Code, Cursor, and CodexInitial release of the Mem0 plugin for Claude Code and Cursor, followed by Codex support:

9 MCP memory tools: add_memory, search_memories, get_memories, get_memory, update_memory, delete_memory, delete_all_memories, delete_entities, list_entities: all via mcp.mem0.ai.
Lifecycle hooks: Automatic memory capture at session start, context compaction, task completion, and session end.
Cloud MCP server: Managed endpoint replaces local MCP and Smithery setup.
Streamable HTTP transport: New MCP transport protocol for real-time streaming.
Codex-specific skill: Dedicated skill in mem0-plugin/skills/mem0-codex for Codex workflows.

Claude Code plugin v0.3.3

Improvements:

Search: The search_memories tool description no longer tells the agent to call it before answering anything that could depend on prior context. It now asks for a search before repeating investigation or when earlier decisions, fixes, commands, or results may help, which reduces unnecessary searches (#7420)
Sidekick: Sidekick searches memories when earlier sessions could help, instead of before every answer (#7420)
Extraction: Repository memory instructions are shorter. They no longer ask for a dedicated memory for each command that failed and was then fixed, and no longer carry separate rules against saving personal preferences or memories that only name the repository, branch, or directory (#7420)
Search skill: /search no longer describes categories as best-effort labels or asks for a retry without the category (#7420)
Packaging: PLUGIN_VERSION bumped to 0.3.3, so the mem0-plugin/<version> wire header and plugin_version telemetry field identify builds with these prompts (#7420)

Claude Code plugin v0.3.2

Improvements:

Telemetry: PLUGIN_VERSION bumped to 0.3.2. The mem0-plugin/<version> wire header and plugin_version telemetry field now reflect the fixes from #7322 through #7358 (#7373)
Telemetry: Events are no longer delivered twice, no longer lose parked events on flush, and now attribute each event to the plugin that produced it (#7323, #7324, #7358)

Changes:

Sidekick: Sidekick is now available only in Claude Code, with Sonnet, worktree isolation, and parent memories.

Claude Code plugin v0.3.1

Changed:

Extracted hook orchestration and memory behavior into the shared Python core; Claude transcript parsing remains in its native adapter. The installed package contains the generated runtime rather than importing files outside its plugin directory.
Preserves the public mem0 name, hook declarations, MCP launch configuration, user configuration, and mem0:sidekick worktree behavior. The manifest and marketplace version advance from 0.3.0 to 0.3.1 so installations can identify the update.

Fixes:

Session-end extraction no longer appends a final answer already captured from the transcript while an earlier extraction was running.
Existing repository memories remain searchable after the shared-ID change; explicit shared-memory deletion also covers the legacy ID. Background workers and control skills consistently use Claude’s data directory.
Receives the shared JSON-secret redaction and nested telemetry filtering fixes. The local search tool exposes query, result count, category, scope, and optional run_id for session-specific recall across all scopes.

#7203

Cursor plugin v0.3.3

Improvements:

Search: The search_memories tool description no longer tells the agent to call it before answering anything that could depend on prior context. It now asks for a search before repeating investigation or when earlier decisions, fixes, commands, or results may help, which reduces unnecessary searches (#7420)
Extraction: Repository memory instructions are shorter. They no longer ask for a dedicated memory for each command that failed and was then fixed, and no longer carry separate rules against saving personal preferences or memories that only name the repository, branch, or directory (#7420)
Search skill: /search no longer describes categories as best-effort labels or asks for a retry without the category (#7420)
Packaging: PLUGIN_VERSION bumped to 0.3.3, so the mem0-plugin/<version> wire header and plugin_version telemetry field identify builds with these prompts (#7420)

Cursor plugin v0.3.2

Improvements:

Telemetry: PLUGIN_VERSION bumped to 0.3.2. The mem0-plugin/<version> wire header and plugin_version telemetry field now reflect the fixes from #7322 through #7358 (#7373)
Telemetry: Events are no longer delivered twice, no longer lose parked events on flush, and now attribute each event to the plugin that produced it (#7323, #7324, #7358)

Changes:

Sidekick: Removes Sidekick and its start/stop hooks. Memory capture, search, and six skills remain available.

Cursor plugin v0.3.1

Changed:

Moves from the legacy shared editor-plugin directory to a native integrations/cursor-plugin/ package with generated Python core and skills, Cursor variables, local MCP configuration, and a native Sidekick.
Translates Cursor conversation/workspace fields, response and summary fields, tool outcomes, and subagent lifecycle events into the shared runtime. The Sidekick searches memory itself because Cursor’s subagent-start response cannot inject parent context.

Fixes:

Adapter errors are logged and exit successfully so a memory failure does not terminate the host hook. Sidekick telemetry reports zero parent-injected context because this host uses self-search.
Repeated response events and Stop/session-end capture do not duplicate the same answer.
Receives shared data-directory handling, background-worker identity, redaction, and legacy-memory retrieval fixes.

#7203

Codex plugin v0.3.3

Improvements:

Search: The search_memories tool description no longer tells the agent to call it before answering anything that could depend on prior context. It now asks for a search before repeating investigation or when earlier decisions, fixes, commands, or results may help, which reduces unnecessary searches (#7420)
Extraction: Repository memory instructions are shorter. They no longer ask for a dedicated memory for each command that failed and was then fixed, and no longer carry separate rules against saving personal preferences or memories that only name the repository, branch, or directory (#7420)
Search skill: /search no longer describes categories as best-effort labels or asks for a retry without the category (#7420)
Packaging: PLUGIN_VERSION bumped to 0.3.3, so the mem0-plugin/<version> wire header and plugin_version telemetry field identify builds with these prompts (#7420)

Codex plugin v0.3.2

Improvements:

Telemetry: PLUGIN_VERSION bumped to 0.3.2. The mem0-plugin/<version> wire header and plugin_version telemetry field now reflect the fixes from #7322 through #7358 (#7373)
Telemetry: Events are no longer delivered twice, no longer lose parked events on flush, and now attribute each event to the plugin that produced it (#7323, #7324, #7358)

Changes:

Sidekick: Renames shared tracking to use subagent terminology. Native subagent memory support remains available.

Codex plugin v0.3.1

Changed:

Moves from the legacy shared editor-plugin directory to a native integrations/codex-plugin/ package, with generated Python core and six skills, a local search MCP server, and native lifecycle hooks.
Resolves MCP repository searches from Codex workspace metadata when supplied. Control skills, hooks, and workers use the same plugin data directory.
Native subagent start/stop hooks supply parent-retrieved memory context and record completions for every native subagent. Named custom agents remain project/user configuration; the plugin does not distribute a named Codex Sidekick.

Fixes:

Receives shared credential redaction, legacy-memory retrieval, background-worker identity, and duplicate-response fixes. Codex tool outcomes use available structured failure indicators; missing outcome information is recorded as unknown.

#7203

OpenCode plugin v0.4.1

Improvements:

Search: The search_memories tool description no longer asks the agent to search proactively or to run several searches for multi-part questions. It now asks for a search before repeating investigation or when earlier decisions, fixes, commands, or results may help, the same wording as the other coding-agent plugins (#7420)
Session context: Removed two system-context lines that told the agent to run 2 parallel searches before responding and 2-4 parallel searches for non-trivial tasks (#7420)
Skills: /mem0-search and /mem0-context-loader make one search_memories call instead of 2 and 2-4 parallel calls. /mem0-context-loader now uses the search skill description from the other coding-agent plugins instead of asking to load at every new task or context switch (#7420)

OpenCode plugin v0.4.0

Changes:

Telemetry: The PostHog source tag changed from the literal "plugin" to OPENCODE_PLUGIN, and project_hash is now salted. Saved PostHog insights filtering on source = "plugin" will stop matching new events; historical data is unaffected (#7322)
Config: A new keyFingerprint key appears in the install-count deduplication logic; installs are now counted once per key rather than on every activation (#7325)

OpenCode plugin v0.3.0

Changed:

Moved the source from integrations/mem0-plugin/.opencode-plugin/ to integrations/opencode-plugin/, retaining the @mem0/opencode-plugin package name and native OpenCode hooks.
Reuses shared conversation preparation, redaction, scoping, and telemetry. Builds a self-contained Bun/ESM dist/index.js and publishes its TypeScript entry declaration.
Global memory tool scope requires the user to enable it in plugin settings first; empty and wildcard identities are rejected.
Retains the seven commands for context loading, search, remember, forget, scope, status, and tour; bundled skills continue loading through OpenCode’s native configuration.

Removed:

Removed auto-Dream consolidation, its gates and state handling, and the Dream and pin skills/commands. Existing configurations and workflows that use these features must be updated.

Builds:

Updated build and publish paths for the relocated source directory; the existing release tag prefix and publishing workflow filename are unchanged.

#7203

OpenCode plugin v0.2.2

Fixes:

Shell-profile API key recovery: When MEM0_API_KEY isn’t set in the process environment, the plugin now falls back to reading it from .zshrc, .bashrc, .zprofile, .bash_profile, or .profile, fixing startup failures on clients (e.g. Desktop) that launch without shell-exported environment variables.

OpenCode plugin v0.2.1

Improvements:

Sharper tool descriptions: Every memory tool (get_memories, get_memory, update_memory, delete_memory, delete_all_memories, delete_entities, list_entities, get_event_status) now states what it does and when to use it. search_memories nudges proactive, multi-hop retrieval and the destructive tools carry explicit irreversible-action warnings.

OpenCode plugin v0.2.0

Breaking Changes:

Native SDK tools: Memory tools are registered through @opencode-ai/plugin and backed by the mem0ai SDK directly. The plugin no longer registers or depends on mcp.mem0.ai.
Skills load in place: The config hook adds opencode-skills/ to skills.paths; skills are no longer copied into user config directories.
Focused skill set: Trimmed to 9 skills: context-loader, dream, forget, status, search, scope, pin, remember, and tour.

New Features:

Memory scopes: Memory operations support project, session, and global scope, with /mem0-scope for changing the default.
Auto-dream: Gated memory consolidation can merge duplicates, drop stale/sensitive entries, and rewrite vague memories.
Expanded telemetry: OpenCode emits the shared plugin.* schema, including prompt, bash-error, pre-compact, session-stop, dream-triggered, and dream-completed events.

Fixes:

Project ID detection: Git remote parsing now handles HTTPS, SSH, and custom SSH host aliases before falling back to the repo root.
Auto-dream visibility: /mem0-status shows which dream gate is blocking consolidation.
Delete-all safety: User-wide deletes require explicit scope="global", regardless of the default scope.

OpenCode plugin v0.1.3

New Features:

File-context injection: Before a file read, the plugin searches for memories referencing that path and injects prior work as system context.
Structured compaction summaries: Session compaction stores session_summary memories with infer=True.
Recent activity timeline: Initial memory loading shows typed memories with age labels instead of bare text.
PostHog telemetry: Anonymous plugin telemetry uses the shared plugin.* schema and never sends memory content, prompts, or API keys.

OpenCode plugin v0.1.2

New Features:

Automatic coding categories: Sets up 17 coding categories at startup, idempotently and in the background.
Global search mode: Adds global_search support across initial load, per-message recall, resume detection, error lookup, and compaction context.
Global search env: Exports MEM0_GLOBAL_SEARCH to child shells via the shell.env hook.

Changes:

Search filters now follow the global_search setting instead of always using project-scoped filters.
Resume and error-pattern searches broadened by removing narrow metadata type filters.

OpenCode plugin v0.1.1

Changes:

Renamed the package to @mem0/opencode-plugin (#5272).
Added the plugin array to the bundled opencode.json (#5271).
Fixed tsconfig, added publishConfig and the bun lockfile (#5273).
Validated the CI/CD publish flow (#5288).

OpenCode plugin v0.1.0

Initial Release:

Native OpenCode plugin: Added @mem0/opencode-plugin with TypeScript hooks for chat messages, tool execution, system-message transforms, session compaction, and shell env injection.
Memory tooling: Added add, search, update, delete, delete-all, entity tools, session-start memory loading, per-prompt semantic search, error-pattern lookup, auto-capture, and compaction capture.
Skills and commands: Bundled 16 OpenCode-native Mem0 skills and slash commands.
Safety: Blocks MEMORY.md writes and redacts secrets before memory storage.

Antigravity plugin v0.3.3

Improvements:

Search: The search_memories tool description no longer tells the agent to call it before answering anything that could depend on prior context. It now asks for a search before repeating investigation or when earlier decisions, fixes, commands, or results may help, which reduces unnecessary searches (#7420)
Extraction: Repository memory instructions are shorter. They no longer ask for a dedicated memory for each command that failed and was then fixed, and no longer carry separate rules against saving personal preferences or memories that only name the repository, branch, or directory (#7420)
Search skill: /search no longer describes categories as best-effort labels or asks for a retry without the category (#7420)
Packaging: PLUGIN_VERSION bumped to 0.3.3, so the mem0-plugin/<version> wire header and plugin_version telemetry field identify builds with these prompts (#7420)

Antigravity plugin v0.3.2

Improvements:

Telemetry: PLUGIN_VERSION bumped to 0.3.2. The mem0-plugin/<version> wire header and plugin_version telemetry field now reflect the fixes from #7322 through #7358 (#7373)
Telemetry: Events are no longer delivered twice, no longer lose parked events on flush, and now attribute each event to the plugin that produced it (#7323, #7324, #7358)

Changes:

Sidekick: Removes Sidekick. Memory capture, search, and six skills remain available.

Antigravity plugin v0.3.1

Changed:

Ships a native package with generated Python core and six skills, a local search MCP server, a Sidekick declaration, and an adapter for PreInvocation, PostToolUse, and Stop.
Normalizes conversation IDs, transcript paths, workspace paths, tool calls, and errors. The adapter captures all completed user/assistant transcript messages incrementally, including later-turn intent, without replaying earlier messages.
Sidekick searches Mem0 itself; this plugin does not provide Claude Code’s worktree isolation.

Fixes and host limitations:

Accepts MEM0_CWD as an explicit workspace fallback when the host omits workspacePaths. Skips capture when neither is available, rather than writing under an unrelated directory.
Documents global MCP registration with agy mcp add when the host does not register the plugin-scoped server. Recall on the initial invocation depends on the host providing a prompt or readable transcript.
Receives the shared data-directory, redaction, legacy-memory retrieval, and duplicate-response fixes.

#7203

Antigravity plugin v0.1.7

Fixes:

Hooks: The mem0-ensure-deps and mem0-session-start hook commands in hooks.json no longer redirect stderr to /dev/null. Both commands still end in || true so a failure can’t block startup, but a broken dependency install or session bootstrap now shows up in the Antigravity hook log instead of failing invisibly (#6948)

Antigravity plugin v0.1.6

Fixes:

Settings loading no longer crashes on a malformed settings.json: If ~/.mem0/settings.json parsed as valid JSON but wasn’t an object, every hook that resolves identity previously raised, not just setup. It now keeps the defaults instead, and setup only reports it “Created” the settings file when it actually wrote one.

Improvements:

Unknown settings key warning: Setup now flags keys in settings.json it doesn’t recognize instead of ignoring them silently.

Removed:

Dropped the doc-search skill’s “openmemory” section now that OpenMemory has been removed from the docs site.

Antigravity plugin v0.1.5

Fixes:

Assistant messages no longer stored as your own: The session-summary hook (fires at the end of every assistant turn) and the post-compaction hook were sending the assistant’s own message to Mem0 tagged role: "user". Because Mem0 extracts facts about the user from each message and uses role to decide who spoke, the assistant’s first-person prose was being saved as the human’s stated preferences: “I recommend we drop Redis” became User prefers dropping Redis entirely. Both hooks now send role: "assistant", so the same session is stored as Assistant recommended....

Existing memories written by the previous versions are not rewritten. If your memories contain preferences you never expressed, delete them; the plugin will not recreate them.

Antigravity plugin v0.1.4

New Features:

Query-driven prefetch: The user-prompt hook now searches Mem0 with the current prompt and injects the top matches, so relevant memories are guaranteed in context instead of relying on the agent to search for them. Opt out with MEM0_PREFETCH=false.

Improvements:

Multi-hop guidance: The decision rubric now tells the agent to run follow-up searches for multi-part or comparative questions instead of stopping after one.

Antigravity plugin v0.1.3

Fixes:

Readable session metadata: Stop-hook session summaries now store files_touched as a real list instead of a double-encoded JSON string, so returned memories show clean file paths.

Improvements:

Reranked injected context: File-context, bash-error lookup, and session-resume recall now rerank by default for higher-quality automatic context. Disable with MEM0_RERANK=0, false, no, or off.

Antigravity plugin v0.1.2

Fixes:

Telemetry attribution: Antigravity is now reported as antigravity instead of being counted as Claude Code.
Per-editor version reporting: Telemetry reads Antigravity’s own plugin.json, so events report the Antigravity plugin version instead of the Claude plugin version.

New Features:

Platform override support: Shared scripts honor MEM0_PLATFORM when an editor needs to pin its telemetry platform explicitly.

Antigravity plugin v0.1.0

Initial Release:

Self-contained plugin: Added Antigravity plugin packaging with plugin.json, mcp_config.json, hooks.json, lifecycle hooks, shared scripts, and shared Mem0 skills.
AGENTS.md support: Uses Antigravity’s contextFileName: "AGENTS.md" convention.
Shared memory layer: Reuses the Mem0 Platform MCP server, shared hook scripts, and 16 slash-command skills.
Lifecycle hooks: Wires session start, user prompt, file read, memory-tool metadata enforcement, bash output, post-tool tracking, and stop summaries.
Install options: Supports degit installs into ~/.gemini/config/plugins/mem0 and local Antigravity plugin installs.
Codex parity: Added missing Codex Write/Edit/MultiEdit blocking and pre-compaction hooks during the shared-plugin restructure.

Kimi Code plugin v0.3.3

Improvements:

Search: The search_memories tool description no longer tells the agent to call it before answering anything that could depend on prior context. It now asks for a search before repeating investigation or when earlier decisions, fixes, commands, or results may help, which reduces unnecessary searches (#7420)
Extraction: Repository memory instructions are shorter. They no longer ask for a dedicated memory for each command that failed and was then fixed, and no longer carry separate rules against saving personal preferences or memories that only name the repository, branch, or directory (#7420)
Search skill: /search no longer describes categories as best-effort labels or asks for a retry without the category (#7420)
Packaging: PLUGIN_VERSION bumped to 0.3.3, so the mem0-plugin/<version> wire header and plugin_version telemetry field identify builds with these prompts (#7420)

Kimi Code plugin v0.3.2

Improvements:

Telemetry: PLUGIN_VERSION bumped to 0.3.2. The mem0-plugin/<version> wire header and plugin_version telemetry field now reflect the fixes from #7322 through #7358 (#7373)
Telemetry: Events are no longer delivered twice, no longer lose parked events on flush, and now attribute each event to the plugin that produced it (#7323, #7324, #7358)

Changes:

Sidekick: Removes Sidekick and its start/stop hooks. Memory capture, recall, and six skills remain available.

Kimi Code plugin v0.3.1

Changed:

Ships a self-contained native package with six skills, nine lifecycle hooks, a local search MCP server, and a native Sidekick declaration. Added a dedicated installation and troubleshooting guide.
Translates Kimi’s session, prompt, tool, compaction, shutdown, and subagent events into the shared Python runtime. Sidekicks receive parent memory context through Kimi’s native lifecycle.

Fixes:

Recovers completed assistant output from Kimi’s indexed v2 wire transcript when Stop events omit the response text. Repeated Sidekick invocations receive distinct run identifiers. Stops without a host ID are left uncorrelated when multiple matching runs are active, preserving their responses without assigning them to the wrong run.
Keeps controls, hooks, MCP, and detached workers on the same Kimi data directory and preserves host identity in background workers.
Receives the shared redaction, legacy-memory retrieval, and duplicate-response fixes.

Host compatibility:

Documents CHOKIDAR_USEPOLLING=1 for the observed macOS watcher issue. Filesystem isolation remains Kimi’s responsibility.

#7203

kimi-plugin v0.1.0

Initial release of the Mem0 plugin for Kimi Code, sharing its scripts, skills, and marketplace listing with the Claude Code / Cursor / Codex / Antigravity plugin family (#6919)New Features:

MCP server: Registers the hosted Mem0 MCP server at https://mcp.mem0.ai/mcp/, authenticated via the MEM0_API_KEY env var as a bearer token.
Lifecycle hooks: Wires SessionStart (loads context through the context-loader skill), UserPromptSubmit, three PreToolUse hooks (blocks direct Write/Edit/MultiEdit to memory files, enforces metadata defaults on Mem0 MCP tool calls, and injects context on file reads), two PostToolUse hooks (post-tool tracking and Bash-output scanning), Stop, and PreCompact.
Hook adapter: kimi_hook_shim.sh normalizes Kimi Code’s hook contract to what the shared hook scripts expect: it resolves the real project directory from the payload’s cwd (Kimi forces the plugin root as the working directory), converts the array-shaped prompt field and the tool_output / tool_input.path field names to the Claude-style shapes the scripts already handle, and translates the plugin-scoped MCP tool name prefix.
Shared policy skill: Bundles the /mem0:policy skill for managing the ## Instructions and ## Agent Instructions sections of mem0.md.

openclaw-mem0 v1.2.1

Improvements:

Search: The memory_search tool description no longer asks the agent to search proactively or to run several searches for multi-part questions. It now asks for a search before repeating investigation or when earlier decisions, fixes, commands, or results may help. Recall strategies (smart, always, manual) are unchanged (#7420)

openclaw-mem0 v1.2.0

Changes:

Config: Added keyFingerprint to the config schema for install-count deduplication; installs are now counted once per key rather than on every activation (#7325)
Telemetry: Events are no longer delivered twice, and the plugin_version field now reflects the plugin that produced the event (#7323, #7324, #7358)

openclaw-mem0 v1.1.0

Changed:

Reuses shared conversation preparation, redaction, and telemetry while retaining OpenClaw’s native memory backend, tools, CLI, and Platform/OSS modes.
Continues to publish a self-contained ESM package under @mem0/openclaw-mem0; the plugin manifest and package version now agree.

Removed:

Removed Dream consolidation: automatic scheduling and locking, openclaw mem0 dream, Dream configuration, the memory-dream skill, and Dream-state public artifacts. Triage, recall, and memory/entity artifacts remain available. Update configurations or integrations that use the removed Dream surface.

Fixes:

openclaw mem0 status handles an unconfigured installation without crashing and directs users to setup.
Removed OpenClaw’s separate 2,000-character extraction cutoff. Selected user and assistant messages retain their full redacted text; recent-message selection, earlier summary selection, and noise filtering still apply.
Telemetry removes sensitive properties recursively and uses the shared failure-safe delivery implementation.

#7203

openclaw-mem0 v1.0.16

Security:

Dependencies: Tightened the undici pnpm override from <6.27.0 → >=6.27.0 <8.0.0 to <7.29.0 → >=7.29.0 <8.0.0, closing a newer CVE range the previous floor didn’t cover, as part of a wider dependency patch sweep across the pnpm workspaces (#6847)

openclaw-mem0 v1.0.15

Improvements:

Onboarding suggestions: The example commands shown by openclaw mem0 config show now suggest gpt-5-mini instead of gpt-4o (#6704)

Security:

Dependencies: Patched high and medium severity dependency vulnerabilities via pnpm.overrides (protobufjs, axios, postcss, mongoose) (#6639)

openclaw-mem0 v1.0.14

Improvements:

Sharper tool descriptions: Every memory tool (memory_search, memory_get, memory_list, memory_add, memory_update, memory_delete) now states what it does and when to use it. memory_search nudges proactive, multi-hop retrieval; memory_add nudges proactive saves; and the delete tools carry explicit irreversible-action warnings.

openclaw-mem0 v1.0.13

Fixes:

Custom categories payload: customCategories is converted into the list shape expected by the Mem0 SDK before add calls (#5345).
Metadata registration: register() exits early in cli-metadata mode, avoiding backend initialization and hook installation during OpenClaw metadata scans (#5383).

Security:

Bumped mem0ai to 3.0.7 and added a uuid override to resolve known dependency advisories (#5460, #5489).

Improvements:

Repo consolidation: OpenClaw moved to integrations/openclaw/; npm provenance now points to the correct subdirectory (#5491).

openclaw-mem0 v1.0.12

Docs:

Agent Mode onboarding: README documents mem0 init --agent --json for autonomous setup without email, OTP, or browser.

Security:

Added overrides for langsmith, picomatch, vite, and @qdrant/js-client-rest.

Dependencies:

Bumped mem0ai to 3.0.3 and refreshed Vitest coverage tooling.

openclaw-mem0 v1.0.11

New Features:

Skills-mode auto-setup: Onboarding enables triage, recall, dream consolidation, and full tool profile automatically.
Memory runtime capability: Exposes runtime.getMemorySearchManager() and resolveMemoryBackendConfig().
Dimension-aware collections: OSS wizard creates a new collection when embedder dimensions change.
Tool docs in skills: Memory triage and dream skills now include full tool reference sections.

Improvements:

Auto-capture and auto-recall default to enabled.
Skills prefer memory_update over delete+add for atomic edits.
Default search threshold lowered to 0.1 for broader recall.
Config writes preserve OpenClaw gateway plugins state.

openclaw-mem0 v1.0.10

Security:

Telemetry identity now uses SHA-256 hashes instead of MD5, and user email is hashed before telemetry submission.

Fixes:

Version-pinned installs are rewritten to updatable specs.
searchThreshold is standardized across docs, README, and manifest.
PLUGIN_VERSION is injected at build time from package.json.

Manifest Compliance:

Removed non-spec manifest fields and moved endpoint declarations to spec-compliant providerEndpoints.

openclaw-mem0 v1.0.9

Security & Compliance:

Manifest security metadata: Added required env var declarations, sensitive config schema fields, data locations, privacy notes, and external endpoint declarations for platform and OSS modes.
Static-analysis cleanup: Tests now use vi.stubEnv / vi.unstubAllEnvs instead of direct environment access in places flagged by plugin scanning.

Tests:

Expanded coverage to 421 tests across 15 files.

openclaw-mem0 v1.0.8

New Features:

OSS onboarding wizard: Guided setup for LLM provider, embedding provider, vector store, and user ID.
Agent-friendly CLI: --json added to all 16 CLI commands.
Non-interactive OSS setup: Added --mode open-source with --oss-llm, --oss-embedder, and --oss-vector.
JSON helpers: Added structured jsonOut, jsonErr, and redactSecrets helpers.

openclaw-mem0 v1.0.7

New Features:

Chat-based setup: Users can configure the Mem0 platform connection inside OpenClaw chat instead of editing config files manually.

Improvements:

V3 SDK compatibility: Bumped the mem0ai dependency to 3.0.1 for Mem0 v3 API support.
Config cleanup: Removed deprecated orgId, projectId, and enableGraph setup paths from docs and CLI prompts.
Noise filtering: Expanded memory-add filtering and improved JSON extraction when model output includes leading text.

openclaw-mem0 v1.0.6

Fixes:

Telemetry identity: Anonymous users now get a persistent per-machine hash instead of the shared anonymous-openclaw fallback.
Telemetry stitching: First authenticated runs emit $identify so anonymous history can attach to the authenticated profile.
CLI event flushing: Short-lived CLI invocations flush queued telemetry before exit.
Email resolution: API-key-only users resolve email lazily through /v1/ping/.

Improvements:

OpenClaw source tag: Provider calls now send source: "OPENCLAW" across add, search, list, recall, CLI, and OSS backend paths.

openclaw-mem0 v1.0.5

Fixes:

Init menu selection: openclaw mem0 init number choices now select the intended option.
OSS pgvector recovery: Warmup errors now reset initialization so failed pgvector clients do not cascade into repeated client.connect() failures.

Removed:

Removed orgId, projectId, and enableGraph from config, CLI, providers, backend, and tools because API keys are project-scoped and graph memory moved to platform-level configuration.

openclaw-mem0 v1.0.4

OpenClaw Plugin: Production-Ready

Interactive init flow: openclaw mem0 init supports email verification and direct API key setup.
Unified tools: memory_add and memory_delete replace legacy memory tools and match Mem0 naming.
CLI commands: Added init, status, config, import, event list, and event status.
Modular architecture: Extracted tools, CLI, filesystem safety, and backend modules.
Security and tests: Added path traversal protection, pinned dependencies, and 329 tests across 10 files.

openclaw-mem0 v1.0.3

Fixes:

Path traversal protection: Added safePath() containment checks for skill and domain-overlay reads.
Noise filter correction: Restored the Post-Compaction pattern after an incorrect rename.

Security:

Pinned mem0ai to exact 2.3.0.

openclaw-mem0 v1.0.2

Fixes:

Env resolution: Removed plugin-side env-var expansion from config because OpenClaw already resolves env references before plugin registration.

openclaw-mem0 v1.0.1

New Features:

CD workflow: Added trusted-publishing release workflow.
Plugin metadata: Added compatibility and build metadata to package.json.
License: Added Apache-2.0 license file.

Fixes:

Dream gate: Fixed cheap-first ordering, session isolation, and completion verification.
Graceful startup: Plugin starts cleanly when no API key is configured.

openclaw-mem0 v1.0.0

New Features:

Skills-based memory architecture: Added the skill loader, memory triage skill, memory dream skill, batched extraction, and skill-aware recall.
Dream gate: Added idle-period memory consolidation.
Enhanced recall: Added recall.ts with improved retrieval and injection logic.
Plugin manifest: Added openclaw.plugin.json and setup helper scripts.

openclaw-mem0 v0.4.1

New Features:

Extraction quality: Improved noise filtering, deduplication, and extraction instructions.

Fixes:

Improved credential, API key, and secret detection during extraction.
Prevented standalone timestamps from being stored as memories.

openclaw-mem0 v0.4.0

New Features:

Trigger filtering: Skips recall and capture for cron, heartbeat, automation, and schedule triggers.
Subagent isolation: Routes ephemeral subagent sessions back to the parent namespace.
Dynamic recall thresholding: Drops candidates below 50% of the top score.
OSS resilience: Adds SQLite init retry and oss.disableHistory.

Fixes:

Lifecycle hooks now use ctx.sessionKey directly, removing a concurrent-session race.

openclaw-mem0 v0.3.1

New Features:

Message filtering pipeline: Multi-stage noise removal before extraction.
Broad recall: Short or new-session prompts trigger secondary broad search.
Temporal anchoring: Extraction instructions include the current date.

Changes:

Extraction now reads the last 20 messages and the monolithic plugin file was split into focused modules.

openclaw-mem0 v0.3.0

Fixes:

Updated mem0ai for the SQLite to better-sqlite3 migration.

openclaw-mem0 v0.2.0

New Features:

Multi-agent isolation: Added per-agent memory isolation via agentId.
Docs: Added guidance for understanding userId.

Fixes:

Migrated platform search to the Mem0 v2 API.

openclaw-mem0 v0.1.2

New Features:

Added a source field for OpenClaw memory entries.

Fixes:

Fixed auto-recall injection and auto-capture message drops.

openclaw-mem0 v0.1.0

Initial Release:

Memory backend: Added Mem0 Cloud and self-hosted OSS modes for OpenClaw.
Auto-recall: Injects relevant memories before each turn.
Auto-capture: Stores facts after each turn.
Config: Added topK, threshold, and apiVersion options.

Pi Agent plugin v0.3.2

Improvements:

Search: The memory policy, mem0_memory tool description, and prompt guidelines no longer ask the agent to search before answering anything that may depend on earlier context or to run several searches per question. They now ask for a search before repeating investigation or when earlier decisions, fixes, commands, or results may help (#7420)
Skills: context-loader makes one search instead of 2-4 parallel searches, and uses the search skill description from the other coding-agent plugins (#7420)
Automatic recall: Injected memories are introduced as “Mem0 found these relevant memories from earlier work in this repository:”, the same heading as the Python plugins. The old heading called them a shallow first pass and told the agent to search again (#7420)

Pi Agent plugin v0.3.1

Improvements:

Telemetry: Events are no longer delivered twice, no longer lose parked events, and now attribute each event to the plugin that produced it. The plugin_version wire field reflects the fixed release (#7323, #7324, #7358)

Pi Agent plugin v0.3.0

Changed:

Reuses shared conversation preparation, memory formatting, project/session/global scope utilities, and telemetry while preserving Pi’s native extension API and @mem0/pi-agent-plugin package name.
Pi loads the built dist/entry.js extension instead of executing source TypeScript from an installed package. Builds also publish the library entry point and declarations.

Removed:

Removed Dream consolidation and pin commands, skills, configuration, types, and exports. The remaining commands are remember, search, forget, tour, scope, and status. Update integrations that import removed APIs or invoke removed commands.

Fixes:

Global memory tool scope requires the user to select /mem0-scope global or configure a global default first. Empty and wildcard identities are rejected.
Memory update and delete accept the mem0:<uuid> and [mem0:<uuid>] citations displayed in tool results, as well as raw IDs.
Shared capture preparation filters conversation roles and redacts content; telemetry removes sensitive keys inside nested structures.

#7203

Pi Agent plugin v0.1.5

Security:

Dependencies: Tightened the undici pnpm override from <6.27.0 → >=6.27.0 <8.0.0 / >=8.0.0 <8.5.0 → >=8.5.0 to <7.29.0 → >=7.29.0 <8.0.0 / >=8.0.0 <8.9.0 → >=8.9.0 <9.0.0, closing a newer CVE range the previous floors didn’t cover, as part of a wider dependency patch sweep across the pnpm workspaces (#6847)

Pi Agent plugin v0.1.4

Security:

Dependencies: Patched high and medium severity dependency vulnerabilities via pnpm.overrides (axios, brace-expansion, postcss, mongoose, protobufjs) (#6639)

Pi Agent plugin v0.1.3

New Features:

Guaranteed retrieval: Memories relevant to each prompt are now prefetched and injected as a <mem0-relevant-memories> block before the agent runs, so context no longer depends on the agent remembering to search. Context injection is on by default; set contextInjection: false to disable.

Improvements:

Multi-hop retrieval: The memory policy and the mem0_memory tool description now push proactive, multi-hop search — several queries with different phrasings instead of stopping after one.
Clearer tool schema: The mem0_memory action and parameter descriptions (action, query, content, memory_id, scope) now explain each action and scope.

Pi Agent plugin v0.1.2

Fixes:

Visible command feedback: /mem0-remember, /mem0-forget, /mem0-pin, and /mem0-scope now render persistent message blocks instead of easy-to-miss status notifications.
Relevance-filtered search: Search, forget, and pin commands now pass threshold, top_k, and rerank to avoid weak unrelated matches.

Improvements:

Richer command output: Commands show scope, query, match counts, and affected memories.
Cleaner dream flow: /mem0-dream keeps the consolidation protocol hidden behind a concise status line.
Configurable relevance floor: searchThreshold defaults to 0.3 in mem0-config.json and is shown in /mem0-status.

Pi Agent plugin v0.1.1

Chores:

Version bump to validate the new pi-agent-plugin-checks.yml and pi-agent-plugin-cd.yml release pipeline.

Pi Agent plugin v0.1.0

Initial Release:

Extension entry point: Registers the mem0_memory tool, 8 slash commands, and auto-capture hooks.
Agent tool: Supports search, add, get-all, delete, and delete-all with scoped filters.
Slash commands: Adds /mem0-remember, /mem0-forget, /mem0-search, /mem0-tour, /mem0-dream, /mem0-pin, /mem0-scope, and /mem0-status.
Auto-capture: Extracts and stores memories from user and assistant messages on agent_end.
Dream consolidation: Merges duplicates, resolves contradictions, and prunes stale entries behind session, time, and memory-count gates.
System prompt injection: Appends MEMORY_POLICY before every agent turn.
Project scoping: Uses git rev-parse --show-toplevel for stable app_id values across monorepos.
Scopes and categories: Supports project, session, and global scopes plus 10 default memory categories.
Agent skills: Bundles context-loader, remember, search, forget, dream, tour, pin, and status skills.
Safety: Confirmation dialogs protect destructive commands, pins preserve memory IDs, full memory UUIDs stay visible, outputs are truncated, all tool actions respect AbortSignal, and shutdown releases dream locks.
Telemetry: Batched PostHog telemetry uses PII-safe error payloads.

mem0-strands v0.1.1

New Features:

Usage telemetry: Anonymous usage events (strands.store.init, strands.store.search, strands.store.add, strands.store.add_messages) ride the Mem0 SDK’s existing PostHog client, unsampled, with no new dependency. Events carry only counts, durations, booleans, and coarse failure kinds: never queries, memory text, message content, entity ids, metadata, or API keys. Opt out with MEM0_TELEMETRY=false (#7110)

mem0-strands v0.1.0

Initial release of mem0-strands, a native MemoryStore that plugs Mem0 into the Strands Agents MemoryManager (#7021)New Features:

Automatic recall and injection: Mem0MemoryStore.search() runs every turn through the MemoryManager, so relevant memories are searched and prepended to the prompt with no explicit tool call required.
Server-side extraction: add_messages() renders raw conversation turns to text and hands them to Mem0’s own extraction pipeline (infer=True), so enabling extraction skips an extra client-side model call to distill facts first.
Verbatim writes: add() stores a single fact exactly as given (infer=False), the sink used by the add_memory tool or a client-side extractor.
Entity scoping: Accepts user_id, agent_id, run_id, and app_id; at least one is required, and mixing the platform-only app_id with a self-hosted config raises at construction instead of failing on the first write.
Hosted or self-hosted: Defaults to the hosted Mem0 Platform via api_key (or $MEM0_API_KEY), or pass a config dict for a self-hosted Mem0 OSS backend.
Non-blocking construction: The underlying Mem0 client is built lazily on first use inside asyncio.to_thread, so API-key validation and OSS embedder/vector-store setup never block the event loop.

Mem0MemoryStore is the automatic-recall store for the MemoryManager. For a model-called tool instead, use the mem0_memory tool from strands-agents-tools; both share the same Mem0 backend and namespace. See Strands Agents for setup.

deepseek-plugin v0.3.2

Improvements:

Search: The search_memory tool description no longer asks the agent to search proactively before answering anything that may depend on earlier context. It now asks for a search before repeating investigation or when earlier decisions, fixes, commands, or results may help (#7420)
Automatic recall: Injected memories are introduced as “Mem0 found these relevant memories from earlier work:”. The old heading called them a shallow first pass and told the agent to search again with mem0_memory, a tool DeepSeek does not have (#7420)

deepseek-plugin v0.3.1

Improvements:

Telemetry: Rebuild with the fixed shared telemetry core from agent-plugin-core. Events are no longer delivered twice, no longer lose parked events on flush, and now attribute each event to the plugin that produced it (#7323, #7324, #7358)

deepseek-plugin v0.3.0

Added:

Automatic recall during system-prompt/assemble, using the latest human prompt and avoiding repeated context injection within a session.
Automatic capture from the durable session/event stream after a completed turn. Interrupted or incomplete turns are not sent through this automatic capture path. autoRecall and autoCapture default to true and can be disabled.

Changed:

Reuses shared lifecycle, redaction, identity, and telemetry utilities while retaining the explicit search_memory and add_memory tools and their per-call agent/session scope. Cross-user userId overrides now require operator opt-in with allowUserOverride: true.
Publishes a self-contained ESM artifact under @mem0/deepseek-plugin; native Harness services and the Mem0 SDK remain external dependencies. Plugin cleanup remains tied to the native Cordis lifecycle.

Host compatibility:

Supports the declared Harness runtime dependencies and documents the macOS watcher workaround. The plugin does not bundle a named Sidekick or provide child filesystem isolation.

#7203

deepseek-plugin v0.1.1

New Features:

Usage telemetry: Anonymous usage events (deepseek.plugin.mounted, deepseek.tool.search_memory, deepseek.tool.add_memory) are batched to PostHog over native fetch and flushed in the background. Events carry only tool names, durations, counts, and coarse failure kinds: never queries, memory text, filters, or API keys. Opt out with MEM0_TELEMETRY=false (#7110)

deepseek-plugin v0.1.0

Initial release of deepseek-plugin, a native DeepSeek Harness (Cordis) plugin that registers Mem0 as two agent-callable tools (#7027)New Features:

search_memory: Recalls facts relevant to a query, with an optional limit (default 10) and per-call userId / agentId / runId scope overrides.
add_memory: Stores a fact for future sessions, tagged source: "DEEPSEEK_HARNESS" for backend attribution; extraction runs asynchronously server-side, so a stored fact may take a moment to become searchable.
Cordis lifecycle: apply(ctx, config) declares inject = ['tools'], so the plugin waits for the harness tool registry to exist, and both tools are registered via ctx.tools.register() so they auto-unregister when the plugin unmounts.
Config: userId is required; apiKey defaults to $MEM0_API_KEY; host optionally points at a dedicated Mem0 Platform base URL (not a switch to self-hosted Mem0 OSS).

Developer preview: auto-capture and auto-recall (memory injected into context automatically, without an explicit tool call) are planned but not yet built. The backend’s KNOWN_EVENT_SOURCES allowlist also needs "DEEPSEEK_HARNESS" added before usage surfaces by name in telemetry rather than bucketing into “OTHERS”. See DeepSeek Harness for setup.

Vercel AI SDK v3.0.3

Improvements:

Client: Inherits the three surface-identity headers (X-Mem0-Source, X-Application, X-Mem0-Client) from the TypeScript SDK bump, so platform calls made through the Vercel AI SDK provider are now correctly attributed (#7326)

Vercel AI SDK v3.0.2

Security:

Dependencies: Tightened the js-yaml pnpm overrides from <3.15.0 → >=3.15.0 <4.0.0 / >=4.0.0 <4.3.0 → >=4.3.0 <5.0.0 to <3.15.1 → >=3.15.1 <4.0.0 / >=4.0.0 <4.3.1 → >=4.3.1 <5.0.0, closing a newer CVE range the previous floors didn’t cover, as part of a wider dependency patch sweep across the pnpm workspaces (#7032)

Vercel AI SDK v3.0.1

Security:

Dependencies: Patched high and medium severity dependency vulnerabilities via pnpm.overrides (brace-expansion, js-yaml) (#6639)

Vercel AI SDK v3.0.0

Major Release: Migrated to Vercel AI SDK v6 (LanguageModelV3 / ProviderV3) and Mem0 v3 API.Breaking Changes:

AI SDK v6: Upgraded from AI SDK v5 (LanguageModelV2) to v6 (LanguageModelV3). Users must upgrade ai to ^6.0.199 and all @ai-sdk/* provider packages to ^3.x (#4741)
Mem0 v3 API: Memory endpoints migrated from /v1/memories/ and /v2/memories/search/ to /v3/memories/add/ and /v3/memories/search/. Entity IDs (user_id, agent_id, run_id) now go inside the filters object for search requests (#4741)
Graph memory removed: All enable_graph, graph prompts, and relation-extraction code removed. Graph memory is now a project-level setting on the Platform (#4741)
Deprecated params removed: org_id, project_id, org_name, project_name, output_format, filter_memories, async_mode, enable_graph, version, api_version removed from Mem0ConfigSettings (#4741)

New Features:

V3 provider contract: specificationVersion: 'v3', supportedUrls property, V3 content array in doGenerate, V3 stream lifecycle events in doStream (#4741)
Mem0 source in responses: Memories are attached as a source in generateText/streamText responses with providerMetadata.mem0.memories for programmatic access (#4741)
Raw search helper: searchMemories() is exported alongside addMemories(), retrieveMemories(), and getMemories() for callers that need raw semantic search results.
Self-hosted API support: mem0ApiKey and host can be passed per request, letting apps override the environment API key and point at a custom Mem0 API base URL.
Provider coverage: Built-in provider selection covers OpenAI, Anthropic, Google/Gemini, Groq, and Cohere through the AI SDK provider packages.
Multimodal memory input: File, image, Markdown, and PDF prompt parts are converted into the Mem0-compatible message shape when adding or searching memories.

Bug Fixes:

Async memory storage: addMemories is now properly awaited: memories no longer silently fail to store (#4741)
Prompt mutation: Prompt array is now cloned before injecting memory context, preventing side effects on the caller’s array (#4741)
Null guard on content: doGenerate guards against null content from upstream providers (#4741)
Stream response: doStream now returns the full LanguageModelV3StreamResult object preserving all V3 fields (#4741)
Response normalization: getMemories and retrieveMemories now handle both array and {results: [...]} envelope responses from the v3 API (#4741)

Vercel AI SDK v2.0.6

Security:

Dependencies: Pinned transitive dependencies via pnpm overrides to remediate high-severity CVEs: glob → ^10.5.0 (CVE-2025-64756), minimatch → ^3.1.3 / ^5.1.8 / ^9.0.7 (CVE-2026-27903, CVE-2026-27904, CVE-2026-26996), picomatch → ^2.3.2 (CVE-2026-33671), rollup → ^4.59.0 (CVE-2026-27606)

Vercel AI SDK v2.0.5

Bug Fix:

Removed unnecessary dependencies to make the package lighter.

Vercel AI SDK v2.0.3 – v2.0.4

New Features:

Added file support for multimodal capabilities with memory context (v2.0.3)

Bug Fix:

Fixed version parameter to use V2 for addition (v2.0.4)

Vercel AI SDK v2.0.2

Bug Fix:

Fixed streaming response in the AI SDK.

Vercel AI SDK v2.0.0 – v2.0.1

New Features:

Migration to AI SDK V5 (v2.0.0)
Added host param to the config (v2.0.1)

Vercel AI SDK v1.0.6

New Features:

Added filter_memories param.

Vercel AI SDK v1.0.5

New Features:

Added support for Google provider.

Vercel AI SDK v1.0.3 – v1.0.4

New Features:

Added support for output_format param (v1.0.4)

Improvements:

Added graceful failure handling when services are down (v1.0.3)

Vercel AI SDK v1.0.1

New Features:

Added support for graph memories.

n8n-nodes-mem0 v0.1.4

Security:

Dependencies: Add js-yaml pnpm overrides (<3.15.1 → >=3.15.1 <4.0.0, >=4.0.0 <4.3.1 → >=4.3.1 <5.0.0) to close a HIGH/CRITICAL severity advisory, as part of a wider dependency patch sweep across the pnpm workspaces (#7032)

n8n-nodes-mem0 v0.1.3

Changes:

License changed to MIT: The published @mem0/n8n-nodes-mem0 package is now MIT (was Apache-2.0). n8n’s Creator Portal requires verified community nodes to be MIT, and the failing license check was the blocker for verification. The rest of the mem0 repo stays Apache-2.0 (#6804)
Themed icons: The node and credential icons now declare { light, dark } variants instead of a single icon, clearing the remaining icon-prefer-themed-variants warnings from the Creator Portal scan. No functional changes (#6804)

n8n-nodes-mem0 v0.1.2

Changes:

Package contact: author.email in the published package is now integrations@mem0.ai (was founders@mem0.ai), so npm and n8n Creator Portal correspondence reaches the integrations team directly. No functional changes (#6791)

n8n-nodes-mem0 v0.1.1

Changes:

Published with npm provenance: Republished through the n8n-nodes-mem0-cd.yml GitHub Actions workflow so the package carries a signed provenance attestation. 0.1.0 was published manually and has none, which blocks submission for n8n Creator Portal verification. No functional changes (#6685)

n8n-nodes-mem0 v0.1.0

Initial release of @mem0/n8n-nodes-mem0, a community node that adds long-term memory to n8n workflows and AI Agents (#6517)New Features:

Memory operations: A single Mem0 node covers Add, Search, Get, Get Many, Update, and Delete on the Memory resource.
AI Agent tool: The node sets usableAsTool, so it can be attached to an n8n AI Agent node and invoked by the agent itself rather than wired into a fixed workflow path.
Scoping: Add, Search, and Get Many accept User ID, Agent ID, App ID, and Run ID, so memories stay partitioned per user, agent, or session.
Add options: Metadata JSON, custom categories, custom instructions, includes/excludes, an infer toggle, and a Wait for Completion switch that polls until the write lands instead of returning immediately.
Pagination: Get Many supports Return All, or explicit Page and Page Size.
Credential: A Mem0 API credential holds the API key plus a configurable base URL, defaulting to https://api.mem0.ai for self-hosted deployments.

Community nodes install from npm, which is a self-hosted n8n feature. See n8n for setup.

Zapier app v0.1.2

Changes:

Connection label: The saved connection now shows the account’s email ({{user_email}}, read from the /v1/ping/ test response) in the Zap editor instead of a static “Mem0” label, so a user with more than one Mem0 connection can tell them apart (#6985)
Action and search copy: Reworded labels and descriptions to address Zapier’s publishing review: Get Memories is now Find Memories by User, Search Memories is now Find Memories, and every description now reads as a third-person statement of what the step does (#6985)
Attribution: Add Memory now tags writes with source: "ZAPIER" in the request body, so usage is attributed to this integration server-side (#6985)

Removed:

Client-side telemetry: Deleted the embedded PostHog telemetry client (telemetry.ts) and its call sites in Add Memory, Get Memories, and Search Memories. Usage attribution now happens server-side via the source: "ZAPIER" tag above instead of a separate fire-and-forget analytics call from inside the published app (#6985)

Security:

Dependencies: Tightened the undici pnpm override to <7.29.0 → >=7.29.0 <8.0.0 / >=8.0.0 <8.9.0 → >=8.9.0 <9.0.0 and added a brace-expansion override (#6847), then added a js-yaml override (<3.15.1 → >=3.15.1 <4.0.0, >=4.0.0 <4.3.1 → >=4.3.1 <5.0.0) closing a further HIGH/CRITICAL severity advisory (#7032)

Zapier app v0.1.1

Bug Fixes:

Add Memory: User ID is now a required field. Mem0 rejects a write that carries no entity ID, so a Zap left blank failed at the API instead of in the editor. If you scoped memories only by Agent ID or Run ID, set a User ID as well (#6790)
Deployment: Add a root index.js that re-exports the compiled app from dist/, and point main at it. Zapier’s Lambda wrapper loads <root>/index.js and ignores the package.json main field, so the deployed app could not resolve its entry point (#6789)

Zapier app v0.1.0

Initial release of the Mem0 Zapier app, built on the Zapier Platform CLI (#6518)New Features:

Actions: Add Memory and Delete Memory.
Searches: Search Memories and Get Memories, usable as lookup steps in any Zap.
Authentication: An API key connection validated against Mem0 the moment it is saved, sent as Authorization: Token <key>, with a configurable base URL for self-hosted deployments.

Bug Fixes:

Add Memory: Raise the poll budget past the real API latency tail, so a slower write is no longer reported as a failure (#6680)

The app deploys to Zapier’s platform rather than npm and is not yet listed in the public App Directory. See Zapier for invite access.