Developer Builds AP Agent With Hindsight Memory That Learns When to Escalate

An accounts-payable desk typically resolves recurring invoice exceptions, such as a monthly steel price adjustment, by checking the contract and approving the charge each time. The author built Ledgerwise, which uses deterministic matching to find discrepancies and an agent to decide them based on how similar exceptions were resolved before. Early versions that simply approved whatever memory suggested failed on a large fuel surcharge that exceeded the contractual cap, prompting a redesign around whether memory is sufficient for a given case. The agent now stores each human decision along with the reason and magnitude, using Hindsight, an open-source agent memory system from Vectorize. The system auto-resolves when precedent is consistent and otherwise escalates with the relevant prior decisions attached.