mnemos

Mnemos is a memory layer for AI coding agents. I built it because Claude Code keeps forgetting conventions, corrections, and decisions between sessions, and that gets old fast.





Most memory tools care about retrieval, how stuff comes back out. I care more about whether anything gets recorded in the first place, because agents almost never bother unless you force them. So mnemos has both, and it ships with a verifier that runs Claude twice on a fixed scenario set (once with mnemos enabled, once disabled) so you can actually see what changes.

The basic shape: you correct your agent for retrying blindly on a 401. Mnemos stores:

{
  "type": "correction",
  "tried": "retry on 401",
  "wrong_because": "401 is auth failure, not transient",
  "fix": "refresh token, then retry once"
}

Next session, that entry surfaces in the prewarm context before the agent gets near that code path again. Three corrections clustered on the same agent, project, and topic get promoted into a skill by the consolidation pass. There's no LLM running inside the memory layer, just clustering and templating over structured records, so the same input always gives you the same output and you can audit the whole pipeline end to end.

mnemos replay regenerates a past session as markdown with everything you've recorded since layered in. Useful for "what would I do differently now" sessions.

Single static Go binary, ~15 MB, runs on Linux, macOS, and Windows. State lives in ~/.mnemos/mnemos.db (SQLite).