Auto memory records the small facts Claude picks up during a session - how to run tests, which commands failed, what you corrected - and surfaces them on the next session.

What it does

Claude writes short notes to a per-project memory file when something seems worth remembering: "tests run with pnpm test --filter web", "the deploy script needs DATABASE_URL exported", "user prefers short commit messages". On the next session, Claude reads those notes into context automatically.

When to use it

Projects where build or deploy steps are non-obvious.
Long-running work that spans many sessions.
Capturing user preferences without manually writing them into CLAUDE.md.
Reducing the "what's the test command?" dance.

Gotchas

The file grows over time. Review and prune periodically or you'll carry stale notes.
Auto memory is local by default. Don't rely on it for team-wide context - use CLAUDE.md for that.
Secrets may slip in if you're not careful. Scan the file before committing it anywhere.

Official docs: https://code.claude.com/docs/en/memory.md#auto-memory