mem0 Release Notes

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

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)

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) …

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) …