MongoGov is governed long-term memory for a customer-support agent, built on MongoDB Atlas. Support runs on repeat questions, and the fix usually already exists, buried in a ticket someone else closed. MongoGov lets an agent learn from closed tickets, but nothing becomes team knowledge until a named human approves it, and every use leaves a receipt. Memory is governed like code review. A learned fix enters as a candidate and stays invisible to the agent until a person other than the proposer promotes it; the agent's attempt to approve its own memory is refused. Every recall passes two separate checks, similarity and a fix's track record, so a close match that failed before is blocked, with the reason written to the record. Memories live in team namespaces, carry a lifecycle (candidate, active, superseded, retired), and can be revised or forgotten, always by a named human. Every step is sealed into a hash-chained receipt before the next one runs, so a process killed mid-ticket resumes from the chain with nothing repeated and nothing lost: disaster recovery for the memory store. Measured today, on the event sandbox: repeat tickets fell from 7.00 to 3.60 tool calls with zero wrong fixes; all three lookalike traps were blocked; a kill test resumed at 10 of 10 kill points; 4 of 4 tampered receipts were caught. Track: Memory You Can Govern.
Team
