Vercel AI SDK Provider (v3.0.0)

Major Release

Migrated to Vercel AI SDK v6 (LanguageModelV3 / ProviderV3) and the Mem0 v3 API.

Breaking Changes

AI SDK v6

Upgraded from AI SDK v5 (LanguageModelV2) to AI SDK v6 (LanguageModelV3).

Users must upgrade:

• ai → ^6.0.199
• All @ai-sdk/* provider packages → ^3.x

Mem0 v3 API

Memory endpoints migrated:

• /v1/memories/ → /v3/memories/add/
• /v2/memories/search → /v3/memories/search/

Entity IDs (user_id, agent_id, run_id) must now be provided inside the filters object for search requests.

Graph Memory Removed

Removed:

• enable_graph
• Graph prompts
• Relation extraction code

Graph memory is now configured as a project-level setting in the Platform.

Deprecated Parameters Removed

The following fields have been removed from Mem0ConfigSettings:

• org_id
• project_id
• org_name
• project_name
• output_format
• filter_memories
• async_mode
• enable_graph
• version
• api_version

New Features

V3 Provider Contract

• specificationVersion: "v3"
• supportedUrls property
• V3 content arrays in doGenerate
• V3 stream lifecycle events in doStream

Mem0 Source in Responses

Memories are attached as a source in generateText and streamText responses.

Programmatic access is available via:

• providerMetadata.mem0.memories

Bug Fixes

Async Memory Storage

addMemories is now properly awaited.

Memories no longer silently fail to store.

Prompt Mutation

Prompt arrays are now cloned before memory context injection.

Prevents side effects on the caller's array.

Null Guard on Content

doGenerate now guards against null content values returned by upstream providers.

Stream Response

doStream now returns the complete LanguageModelV3StreamResult object.

Preserves all V3 fields.

Response Normalization

getMemories and retrieveMemories now support both response formats:

[...]
{ results: [...] }

Original source

Mem0 Python SDK (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_breakdown dict with semantic, keyword (normalized BM25), entity_boost, and temporal_boost signals 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)

Original source

Mem0 Node SDK (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)

Original source

Mem0 Node SDK (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)

Original source

Mem0 Node CLI (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)

Original source

Mem0 OpenCode Plugin (v0.1.2)

Automatic Coding Categories

Categories are now configured automatically in the background when the plugin starts — no manual setup needed.

17 development-focused categories installed on the Mem0 project at startup (architecture_decisions, api_design, security, debugging_notes, etc.)

Fully idempotent: SHA-256 fingerprints of the category list and API key gate re-application (~/.mem0/categories_setup.json)

Never blocks session initialization — runs as a background promise

/mem0:onboard Step 5 now verifies categories exist rather than prompting the user to install them (#5300)

Global Search Mode

New global_search toggle for cross-user, cross-project memory access.

Set global_search: true in ~/.mem0/settings.json to search all memories across all users and projects

Searches use {"OR": [{"user_id": "*"}]} filters; writes still scope to the current user_id and app_id

Applies to all plugin search paths: initial load, per-message recall, resume detection, error-pattern lookup, and compaction context

MEM0_GLOBAL_SEARCH env var exported to child shells via the shell.env hook (#5300)

/mem0:switch-project Expanded

Added --global flag to enable global search (/mem0:switch-project --global)

Added --no-global flag to disable it and return to per-project scoping

Persists to ~/.mem0/settings.json — no manual config editing needed (#5300)

Search Filter Changes

All search paths now dynamically construct filters based on the global_search setting

Resume and error-pattern searches no longer include metadata.type sub-filters (session_state, decision, anti_pattern, bug_fix), broadening recall

System context message updated to inform the model when global search is active (#5300)

Original source

Mem0 Python SDK (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)

Original source

Mem0 Platform v3 can now interpret time-aware memories and queries so assistants retrieve the right information for questions about the past, upcoming plans, and current state.

Time-aware search intent

• Queries like last week, upcoming, right now, and as of March 2025 return contextually appropriate results automatically
• Enabled by default — No per-request toggle required for v3 writes or searches
• Anchored relative queries — reference_date anchors relative search phrases for tests, backfills, and reproducible demos
• Normal response shape — Temporal reasoning affects ranking while preserving existing client response patterns

See Temporal Reasoning for usage details.

Original source