How an AP Agent Uses Selective Memory to Flag Fraud and New Vendor Risk

A developer building an accounts payable AI agent deliberately designed it to return zero memories for first-time vendors, preventing the system from borrowing trust from unrelated suppliers via semantic search. The agent follows a seven-step pipeline — from invoice matching to decision — using Hindsight, an open-source TypeScript memory client, with recall and retain calls bracketing each decision. For known vendors, recalled payment history helps flag anomalies like sudden bank-detail changes, with the agent citing specific past transactions to guide human reviewers toward verified contacts rather than invoice-supplied ones. New vendors are automatically routed to human review by default, while returning vendors benefit from stored resolution patterns, such as a recurring invoice discrepancy that was previously resolved by requesting a corrected invoice. The developer notes that parsed numeric data like payment counts should be stored as structured metadata rather than extracted from prose, flagging it as a known fragility in the current design.