Rule
Capture what the session learned, and re-inject it at the next plan. These are two halves of one loop, and they are worthless separately: a lesson nobody writes down is gone, and a lesson nobody reloads was never a lesson — it was a diary entry.
The discipline is the loop, not the archive. Capture decisions, traps, dead ends, and the commands that mattered; then make the next planning step actually read them, so a recorded mistake biases the next agent away from it. Whether the accumulated record makes each session measurably better than the last is a measured, still-unproven hypothesis — the loop is what you enforce either way, because a closed loop is the only configuration under which that bet can pay at all.
What AgentOps Enforces
- Capture decisions, traps, failed approaches, and load-bearing commands at session close.
- Separate durable lessons from session noise; promote the durable ones into repo-owned files.
- Re-inject relevant lessons at the next plan — retrieval at task start, not archaeology after failure.
- Promote repeated lessons (and recurring dead ends) into stronger rules, checks, or gates.
- Decay or revise stale knowledge when the codebase changes.
Failure Signal
- The next session repeats the same investigation.
- The knowledge base grows but agents keep starting cold.
- The same dead end appears in three different sessions.
- Learnings are written for humans but never loaded by agents.
Done Looks Like
A future agent avoids at least one mistake because this session's lesson was captured — and actually loaded when planning began.