Unsay

agent memory that goes back and un-says what it told you

1 A patient asks

Answered from live openFDA recall data. Every claim the agent reads is recorded with the exact version it read.

2 The world changes

You play the FDA here: this step is staged so the scenario can be run on demand, and the new version is filed under demo:escalation, not under the FDA's name. The lot, the drug and the Class II above are real openFDA records. Publishing retracts the claim on file and asserts a new version under the same key, in one transaction. Nothing already read is mutated.

3 Retroactive repair

Every answer still standing that leaned on a claim version we no longer believe. This is a join across decision → provenance → claim. A vector store cannot ask it: it has no record of which version an answer consumed.

4 Corrections, by name

Each written in the same transaction as the correction that justifies it, keyed so a replayed sweep cannot send a second time.

The sweep is destructive by design. Reset restores the twelve standing answers so the next visitor sees the story unspent.
Nothing corrected yet. Run steps 1 to 3.