How a deliberate memory gap makes this AI accounts payable agent more secure

A developer building LedgerMind, an AI-powered accounts payable agent, intentionally designed it to return zero memories for first-time vendors rather than pulling similar vendors' data as a proxy. The system uses an open-source memory tool called Hindsight to recall and retain vendor payment history, but only queries it when a confirmed payment record exists in the vendor master. This prevents semantic search from fabricating trust by surfacing lookalike vendors — a vulnerability that could be exploited in bank-detail fraud. For known vendors, recalled history helps flag anomalies like sudden account changes, with the agent citing specific past payment counts and verified contacts. New vendors are automatically routed to human review, treating the absence of history as a deliberate safety signal rather than a system gap.