I Published an AI Memory Result. Then Real Retrieval Broke Everything.

A developer turned a private AI memory test into a public demo for evaluating whether agents retrieve the right memory and the one with authority to act.

The system uses layered memory stores, adversarial cases, deterministic evaluators, and a claim ledger to test memory judgment more rigorously.

Early results favored layered memory over summary-only memory, but adding retrieval exposed failures where similar memories overrode the correct policy or jurisdiction memory.

The project now emphasizes authority-based arbitration: the most relevant passage is not always the one that should control the action.