mnem-mcp

An MCP server that exposes a Mnemosyne store to an agent: its memory as tools and resources, so the agent records what it learns, and you can trace a wrong belief later.

pip install mnem-mcp        # not published yet; build from the repo
mnem-mcp --store ./agent-memory  # speaks MCP over stdio

Tools

remember, revise, remember_many, forget, recall, recall_at, history, why, when_did, whats_new. With --tools all, also branch, switch and merge (for a harness, not usually the model).

Every write tool is one commit. Provenance travels on remember as source / step / observation.

Resources

mnem://memory, mnem://memory/{id}, mnem://log, mnem://commit/{id}, so a client keeps the current memory in context without a tool call every turn.

How it binds

One store per server process, from --store or $MNEM_STORE. The server holds no state between calls (the 2026-07-28 MCP transport is stateless); it opens the store fresh each time. Concurrent writes serialise and retry; a persistent loser returns the conflict code.

See ADR-0016.