Skip to content

Submission summary

Memento Emptor is middleware a merchant runs beside its PayPal app. PayPal’s refunds, disputes and captures flow in through verified webhooks. Any shopping agent the merchant runs calls it over MCP before it recommends. It learns why each purchase came back, so agents stop repeating what the shopper already paid for. The layer never moves money.

  • A payment outcome is not a cause. The outcome, the item and the reason are kept apart. Only a matching reason moves a belief.
  • Recall never refuses. It answers in a fixed order: the shopper’s own correction, their own fit evidence, fit pooled across shoppers, then what they said in general. Every answer cites its evidence.
  • The shopper’s correction wins at once, and a second agent that never saw the conversation honours it.
  • Right on day one for a new shopper. In round 2’s new-to-brand case, only the full layer got the size right: 3 of 4 by the committed scores and 4 of 4 by hand, against 0 of 4 for every other setup. We built that case and the pool is seeded, so it shows the mechanism, not real-world accuracy. See the comparison.

PayPal (sandbox, Agent Toolkit, webhooks) and AI (Jev for reasons, Claude Sonnet for the agents), with the sponsor tools AG Grid, Channel3, Elastic, Astropods and Render. See Sponsors for how each was used, and our feedback.

The full form text is in docs/submission.md.