Documentation
Pull request reviews
The reviewer reads a diff with the project's memory beside it — rules, risks, past failures — and posts findings that cite the memory they come from. Every finding you accept or dismiss tunes the memory behind it.
Turn it on
The Reviews screen in the app is a four-step checklist per repo: an AI key on the project (your own Anthropic, OpenAI or Google key), GitHub connected, the reviewer switched on for the repo, and the CI workflow in place. Each step says what is missing and fixes it inline. The reviewer is included from the Small team plan.
Run it in CI
memmo ci # writes .github/workflows/memmo-review.yml
gh secret set MEMMO_TOKEN # a project token from the app (Connect → tokens)The workflow runs memmo review --post on every pull request and posts a review with inline comments. It skips branches named memmo/*, so the rule-files PRs never review themselves. You can also review any open PR on demand from the app.
Decide what counts
Per repo, the policy says which severities to show (blocker, warning, nit, praise) and when the bot approves or requests changes:
- Approve never, when there are no blockers, no warnings, or no findings at all — with an optional cap on warnings.
- Request changes never, on blockers, on blockers or too many warnings, or on any warning.
- Grounded only counts just the findings that cite a memory, so the verdict rests on what the team actually decided.
The review body is verdict first, then counts, then findings grouped by severity. Instructions and skills on the same screen tune what the reviewer looks for.
Close the loop
Mark findings Useful or Noise in the app, or react 👍 / 👎 on the GitHub comments and run memmo review-sync. Useful findings raise the confidence of the memory they cite; noise lowers it. Past reviews stay on the Reviews screen.