MxMMxMMxMMxMMxMMxMMxM

A coding-agent session forgets everything when it ends. The next one rediscovers the same root cause, walks into the same dead end and asks the same questions, because what the last session learnt sits in a transcript nobody reads again.

MxM (Model Experience Memory) keeps that experience. When a session ends it stores each exchange, your prompt and the agent’s reply, in one file per repository. The next session starts with the most recent exchanges, and each prompt you send gets the past exchanges that clearly stand out for it, recalled by meaning with an encoder that runs on your CPU. An unrelated prompt gets nothing: being the closest match isn’t enough, a past exchange has to stand out from the rest of the project’s memory before it’s shown.

After a session ends, a background worker distils it into short facts — decisions and their reasons, root causes, dead ends, conventions — and checks each one against the exchanges it came from before the next session sees it. Facts are never edited in place; retracting one keeps the original on the record.

Legible by design: every block the agent is shown is journaled, and mxm review lets you mark each one useful or noise. It never edits your agent’s settings, it loads its hooks only for the session you start through mxm claude, and if anything fails the session simply runs without memory. The repository also holds the pre-registered research behind it, negative results included.