search_memory is designed to be called ahead of every reply, including trivial ones, and it returns preferences, skills and tool memory under separate limits so one category cannot crowd out the others. Correction is deliberately two-sided: memories are deleted by ID, and feedback records what was wrong, so a bad memory leaves a trace instead of vanishing. Knowledge bases sit apart from memory — documents go in from raw content, a URL or an absolute path, and retrieval can be scoped to particular bases.
A memory service for assistants — dialogue history, retrieved memories, a consolidated user profile, and knowledge bases holding documents.
- Dialogue history saved after an answer, and new information recorded as memory when the user asks for it.
- Retrieval ahead of an answer, with separate limits for memories, preferences, skills and tool memory, a relevance setting, filters, and scoping to chosen knowledge bases.
- A consolidated user profile covering facts, preferences and tool experiences.
- Memory deletion by ID, and feedback that corrects or updates what is already stored.
- Named knowledge bases created and removed, with documents added from content, a URL or an absolute file path, then fetched or deleted by ID.
A MemOS account: the API key in MEMOS_API_KEY, the user it belongs to in MEMOS_USER_ID, and the channel in MEMOS_CHANNEL.
One command plus a key — npx @memtensor/memos-api-mcp, then supply credentials