Researchers have built a memory system for AI agents that refuses to answer rather than repeat a fact it can no longer verify.
A new arXiv paper describes Governed Persistent Memory (GPM), which tracks whether a stored memory has since been contradicted, retracted, deleted, or gone stale, and blocks any answer built on a record in that state instead of surfacing it anyway. On a frozen 3,600-case benchmark, GPM matched every correct outcome, while the best of three simpler baseline policies matched only 1,800 of 3,600 cases and produced bad, unmatched answers on half of the violation cases. In a separate end-to-end service evaluation, the governed system was correct on all 2,400 of 2,400 query clusters, versus 600 of 2,400 for an ungoverned local Qwen2.5-7B model handling the same requests, fixing every one of the baseline's 1,800 failures without adding new errors. A repeat run across English- and Chinese-language command sets matched the same 2,400-of-2,400 score, and separate formal-verification testing spanning roughly 2 million query states and 100,000 execution traces found no contract violations.
Long-running AI agents are only as trustworthy as their memory, and most retrieval setups will happily surface stale or retracted facts if the text resembles a match. GPM's fail-closed design - answer nothing rather than answer wrong - is a real departure from retrieval-augmented systems built to maximize recall, not caution.
These numbers come from a benchmark the same team designed and hash-froze in advance, not a messy real-world deployment, so the harder test is whether "refuse by default" survives contact with agents juggling contradictory data nobody flagged ahead of time.