← Explore
What happens when stored knowledge changes?

explorer

What happens when stored knowledge changes?

Append-only history: corrections add, they never silently overwrite

Agent memory changes constantly: facts get corrected, plans get revised, decisions get reversed. Most systems handle this by overwriting, which destroys the very thing a governance layer needs, namely the record of what was believed at the time an action was taken.

CueCrux takes the opposite approach. The store is an append-only spine. When knowledge changes, the daemon appends a new entry that supersedes the old one; it does not mutate history. Each mutation is signed with a CROWN receipt (Ed25519) and hash-chained (BLAKE3) to everything before it, so the sequence of changes is itself tamper-evident. If an agent acted on stale information last Tuesday, you can reconstruct exactly what the store said last Tuesday.

Checkpoints and sessions build on this. A checkpoint pins a known-good state you can return to or hand to another agent; a session records the working context around a run. Nothing about resuming or branching requires rewriting the past.

Deletion is the one deliberate exception, and it is engineered rather than improvised. Scoped forget removes data on request, with a dry-run mode that shows you precisely what would go before anything does, which is how CueCrux implements GDPR Article 17 erasure. The forget operation is itself receipted, so you can later prove both that the data existed and that it was properly removed.

The result is a store you can argue from. When someone asks why an agent did something, the answer is not a guess about what the memory probably contained; it is a signed, replayable record of what it did contain.