Rule
Retrieve prior learnings — positive and negative — when future work needs them.
This is the read half of the knowledge loop. Factor IX captures the lesson; this factor is what makes capture pay off — extraction without retrieval is a write-only archive no agent ever reads. Failures ride the same flywheel: a recorded dead end is negative knowledge that biases the next agent away from a proven-bad branch, which is often worth more than one more positive pattern.
What AgentOps Enforces
- Search the repo-owned corpus at task boundaries.
- Load only relevant prior lessons into context — including the failures worth avoiding.
- Promote repeated lessons (and recurring dead ends) into stronger rules, skills, or gates.
- Decay or revise stale knowledge when the codebase changes.
Failure Signal
- The knowledge base grows but agents keep starting cold.
- The same dead end appears in multiple sessions.
- Stale advice overrides current repo state.
- Learnings are written for humans but never retrieved by agents.
Done Looks Like
Later sessions start with the useful parts of earlier sessions, and old knowledge loses authority when it stops matching reality.