Mnemosyne Memory (official)

This skill teaches an AI agent to use Mnemosyne, an MIT-licensed, local-first memory engine, through its official MCP server (npm package mnemosy-ai 2.0.0). Memory lives in one SQLite file on your machine. Each memory keeps the source it came from, so the agent can show where a fact, preference, decision or procedure was learned.

What the skill covers:

• Setup. Configure the server in Claude Code (.mcp.json or claude mcp add) with a pinned npx command, plus the project README's Cursor entry (.cursor/mcp.json), which we did not run in Cursor.
• Verification script. It checks Node.js, the CLI version, and a store-and-context round trip in a throwaway database, then prints PASS or FAIL. It runs against an existing install and never downloads anything.
• Session workflow.
  • Resume a handoff.
  • Compile cited context under a token budget.
  • Store new memories with their source.
  • Correct a memory that turned out to be wrong. Memories that declared it as a dependency are marked invalidated and stop being offered as current guidance.
  • Checkpoint before stopping.
• Tool reference. All 17 MCP tools: 16 by default, plus memory_forget when you launch with --allow-destructive. Each entry gives the argument types, an example call and what the tool returns.
• Limits of the model's authority. It cannot switch workspace, raise its own trust or report its own skill trials. It cannot forget unless you allow that at launch.
• Data handling. The database file is mode 0600 in a 0700 directory and stored as plaintext, so keep it out of Git. The five most common setup problems are covered.

The skill reproduces the project's published measurements together with their misses:

• 409 of 500 on LongMemEval-S, measured with reranked retrieval and an LLM reader and judge
• a reranker gain on LoCoMo
• a matched comparison with Mem0 OSS that prints the cases where Mem0 wins
• two release targets that were not met

The default setup in this skill uses lexical retrieval and no reader model, so it does not reproduce the answer-accuracy figure on its own. Evaluate it on your own workload before relying on it.

No API key, database server or Docker is needed, and the default launch makes no model calls.