Evidence both ways, for every claim.
A PhD on how AI agents should forget, run from one bundle: sources, reading notes, claims, experiments and paper drafts. The librarian agent keeps the evidence linked; Nora decides what is true.
- 1 · @noraHands over a PDFShe asks Claude to read Lost in the Middle. Claude reads the PDF, not Korpus.
- 2 · agent/librarianWrites the note, links evidenceAdds the source row and a reading note, and links it as contradicting her claim.
- 3 · @nora mentionedA claim loses its footingThe claim now shows evidence both ways. The librarian cannot edit claims, so it asks.
- 4 · @noraShe revises the claimRev 2 narrows it to eight memories. Her draft cites it; rev 1 stays in history.
Give your agents a memory your team can read.
Korpus is coming soon. For early access, a demo or running it in your own AWS account, write to us.
Get in touch