verify-loop: audit memory one claim at a time
Works a queue of uncertain memory claims, exhausting every source family before bringing the human exactly one question, with the evidence attached.
What it does
An interactive verification loop for the agent's stored knowledge: take one medium-confidence claim, exhaust every source family that could settle it, run the question pre-flight, and only then bring the human the next single question. It powers the Verify lane pattern in this inventory.
Trigger
A verification round over stored claims, run as a standing practice in its own chat.
What you need
- A store of claims with confidence levels, so the uncertain ones form a queue.
- Source families to check before asking: records, messages, files, and the human only last.
- A human willing to rule one question at a time.
Setup sketch
The caveat is the design: the queue is never empty for long, because new claims keep arriving. The skill works precisely because it refuses to declare the data clean while a residue remains.
Deep dive: Memory that holds →