Give the next agent
a lesson you have tested.
Keep reviewed guidance close to its evidence. Deliver it to the devices that run your sessions, then inspect use, guard results, and assessed recurrence.
| Question | Evidence to show |
|---|---|
| Was it approved? | The reviewed package and its exact version. |
| Who received it? | A delivery receipt for each device that installed it. |
| Was it used? | Observed use, or an explicit unknown. |
| Did it help? | Assessed recurrences, matching outcomes, and how much of the window was captured. |
Keep useful guidance. Retire what does not help.
Review conflicting lessons, false positives, and recurring failures. Revise or revoke the package while preserving its history.
Effectiveness is not yet known. RepoOps updates reviewed guidance and controls. It does not retrain the assistant's underlying model.
What happened in this change?
Follow an example from the instruction to the verified fix, then install RepoOps and open your own.
Implementation details & evidence
Make every resolution improve the next attempt.
RepoOps memory is a wiki where every claim carries a citation, a CI check, and a receipt. Six ways it compounds instead of resetting, for you, your team, and the community.
Why you can trust it
Measured, gated, and yours
Repository knowledge
Descriptive, experiential, proactive, and how it compounds
Verified Memory
A generated wiki where every claim cites its source files, a strict CI gate checks those citations against HEAD on every commit, and each claim renders a memory receipt: source file, line, last-verified commit, CI timestamp. We run the same audit on our own brain in public, failing rows included, at repoops.ai/verified.
- Outcome: a page tells you it is fresh as of a specific commit, so you never trust a stale doc.
- Put it to work: open the Wiki tab and read the receipts before you rely on a page.
cited / CI-checked / memory receipts
Experiential memory
Lessons, errors, decisions, and causal links, learned from what actually happened when your agents ran and enforced back into every agent session as pre-merge rules.
- Outcome: the review check compares a diff against your active lessons and fails it when a known mistake comes back.
- Put it to work: let RepoOps write the lesson; it reads on the next run automatically.
Proactive personal brain
Connectors can fill your personal brain on a schedule from email, docs, the web, X, and Slack, and session briefings deliver the relevant piece where you already work.
- Outcome: the context you need arrives at session start, ranked by what you have open.
- Put it to work: connect a source and set your priorities on the personal-brain page.
connectors / briefings / priorities
Lessons write back
The top-ranked fixes land in CLAUDE.md automatically, ordered by how many recurrences each has prevented and bounded so the manual stays readable, with the full set on the Lessons tab. A recurring fix becomes a proposed prevention rule. Your accumulated knowledge is plain files you own, not a database you rent.
- Outcome: future sessions receive the reviewed lesson; actual recurrence and guard catches show whether it prevented the mistake.
- Put it to work: keep your rules in the brain and let new sessions inherit them.
Team brain
Turn sharing on and a member's lessons pool into one shared brain, each carrying the repo it came from and a count of how many bound machines hit it. Sharing is off until each machine enables it, and a lesson that touches a watchlisted sensitive path stays local. A new teammate starts from what the team has already shared, not a blank page.
- Outcome: onboarding a developer or a repo starts from the conventions your team already earned.
- Put it to work: share a slice, and the team wiki synthesizes cross-member topic pages from it.
Brain marketplace
Publish a signed pack of your own lessons to your repo's marketplace view, or install a catalog entry as a staged proposal you review and approve before it lands. Publishing is local: the pack is redacted, signed with an ed25519 key, and listed on your own machine, and nothing is uploaded.
- Outcome: an install is a staged proposal you read and approve, never a silent write into your repo.
- Put it to work: install a catalog entry and approving records it for this repo; applying it is a separate step you take by hand.
publish / install / signature-verified
Ready to start? Install RepoOps and the first run reads your AI coding history and seeds the brain from what your repos already say. Your first lesson lands when your work produces one, a defect, an incident, or a correction, and RepoOps surfaces it when it does. Connect a source and set priorities to grow the proactive side, and read the Wiki tab's verified badges to trust the descriptive side.