MCP tools
The Model Context Protocol is the standard way for agents to discover and call external capabilities. Mnemosyne implements it natively — the same package that powers our REST API also speaks MCP over stdio.
Point any MCP host at Mnemosyne and your agent gets long-term memory with zero glue code.
The tool surface
Mnemosyne exposes 26 MCP tools, covering every cognitive primitive. They are all live in v3-alpha. The handful with a dedicated page are linked; the rest are listed for completeness.
Core recall & write
- memory_recall — search memory for relevant knowledge
- memory_remember — store a new memory
- memory_pin — keep a memory hot and exempt from decay
- memory_forget — retire a memory
memory_search— structured/filtered search across primitives- memory_timeline — chronological view of memories
Knowledge primitives
- memory_entity — manage entities in the knowledge graph
memory_relate— link memories and entitiesmemory_decide— record a decision with its rationale- memory_reasoning — CoT cache: resolve, record, feedback
- memory_strategy — procedural “when X, do Y” patterns
memory_workflow— multi-step procedural memorymemory_skill— reusable learned skillsmemory_hyperedge— n-ary relations across multiple memories
Sessions, episodes, tasks & events
- memory_session — manage a working session (pin, scratch, collapse)
memory_episode_start— open an episodememory_episode_end— close an episode and extract knowledge- memory_task — create, update, delegate, complete tasks
- memory_event — record immutable occurrences
Cognition & meta-memory
memory_status— system/memory health snapshot- memory_introspect — self-model: coverage, gaps, strengths
memory_anticipate— predict what’s likely to be needed nextmemory_consolidate— trigger consolidation / replaymemory_compose— assemble context from multiple memories- memory_self_edit — agent edits its own memory (with mandatory reason)
memory_remind_when— schedule a future reminder on a condition
Shipped in v3-alpha (Beta). Interfaces may still change before 3.0 stable.
Host integrations
How it talks
The MCP server is stdio-only. The host boots it as a child process and
communicates over stdin/stdout — it binds no port and needs no network
listener of its own. (The MCP server reaches your Mnemosyne instance over
HTTP using MNEMO_URL; that’s a client call, not a port it opens.)
Why MCP instead of “import an SDK”
MCP makes the memory tool discoverable. The agent’s system prompt
reads "You have these tools: …" and decides on its own when to call
memory_recall or memory_remember. You don’t have to write any code
that says “if the message mentions X, query memory”. The agent figures it
out.
If you want lower-level control, the TypeScript or Python SDKs are still available.