Memory a person can overrule
Agents learn from lessons. We built the brain so every lesson is checked, can be revoked, and reaches an agent only after a person publishes it.
An agent that learns from its mistakes can also learn the wrong thing, or keep a rule long after the code it describes is gone. So we built memory with a person in charge of it.
The brain
Each repo has a brain: its lessons, decisions, past errors and patterns, kept as plain files. Agents read it through an MCP server and through managed blocks in CLAUDE.md and AGENTS.md. One view shows the brain across every repo. Promoting a lesson to every product is built, and the first promotion waits on connecting our team to PromptCompute.
Lessons that check themselves
- A lesson that names code which no longer exists is held back.
- When two lessons disagree, both are kept for a person to settle.
- A lesson can name the check that enforces it, and we count the lessons that have none.
- Before a lesson goes to the other products, it is replayed against past failures.
Corrections that change the record
In the desktop app, saying “that’s wrong” contradicts the claim, sends every lesson that cites it back to review, marks change-risk cards stale and logs an event on the task. When it disputes the cause, it also counts as a wrong label. On repoops.ai it is a wrong label on the case. In the app you can undo it for ten minutes and revoke it after that. Nothing is deleted: the history keeps both.
Questions an agent asks
Agents ask questions while they work. A decision judge answers each one by citing a decision or lesson, sends it back, or holds it for a person on the Decisions tab. It learns from answers people overturned, pull requests closed without merging, and reverts.
