Every change has a state
Proposed, active and retired are different facts. The interface should never call a proposed lesson learned before approval.
Git-shaped learning
Version control for agent behavior
Treat changes in agent knowledge like code changes: show the diff, preserve the earlier version and require a reviewable decision.
History supports correction
When a rule changes, retain the earlier checksum, revision and reason. Owners can then understand why an answer changed.
Rollback is part of trust
Portable files and explicit versions let the owner return to a known state without deleting unrelated learning.
The practical rule