Save a durable fact about this project so future sessions, in any AI tool, know it. Use this, not a built-in memory or notes file, whenever the user asks you to remember something about the project.

Good facts: conventions ("we use pnpm, not npm"), setup gotchas ("tests need DOCKER_HOST set"),
decisions and their reasons. Not: temporary task details or anything in the code already.
files: repo-relative paths this fact depends on (e.g. ["docker-compose.yml"]), so the memory
can be flagged when those files change.

Search this project's shared memory by meaning. Call it at the start of every task, and whenever project-specific knowledge (setup, conventions, past decisions) might matter. Only current (active) memories are returned; outdated or replaced ones are skipped.

Show where a memory came from (which tool and session saved it, which files it's linked to) and everything that has happened to it since. Use it when a memory looks wrong or surprising.

Mark a memory as no longer true, with a short reason. It stops showing up in recall but keeps its history, so you can still see what it said and why it was retired.

Retire an older memory because a newer one replaces it (e.g. "we use Postgres" replaced by "we migrated to MySQL"). The old memory stops showing up in recall but keeps its history.

Check whether any memory's linked files have changed since it was saved, and retire those memories. recall already does this automatically; call it directly after big changes.

Mark a memory as still true (e.g. after it was retired because a linked file changed, but the fact itself didn't). Re-links it to the current version of its files.

Write this project's current memories into the managed section of AGENTS.md / CLAUDE.md, so they're committed with the code: teammates and tools without Mnemolith see them, and changes show up in reviews. Call it when the user asks, or after a batch of new memories.

List pairs of memories in this project that may contradict each other and need a decision.

Settle a possible conflict. replaces_old=True: the new memory replaces the old one (the old one is retired). replaces_old=False: both are true, keep both. Ask the user if you're not sure.