GovMem cuts AI agent memory errors by 93% using dependency checks

A new paper from researchers Yijiashun Qi, Xiang Xu, and Yuxuan Li tackles a critical safety challenge in long-lived language agents: when should they refuse to write memory? The problem is that repeated observations across agents are not necessarily independent — the same claim may be copied from a shared source, induced by a shared prompt, or stale under a new environment. The authors introduce GovMem, a conservative diagnostic reference policy that estimates dependency-aware support, retrieves counterevidence, assigns scope, and outputs one of three decisions: promote, reject, or needs-review. In controlled synthetic stress tests, GovMem reduced false promotion from 0.597 to 0.040 (a 93% reduction) while maintaining a 0.960 recall, though with an explicit review burden.

In a project-internal real-trace subset of 120 human-labeled candidates spanning 79 recorded traces and project reports, dependency-aware promotion reduced false promotion from 0.371 (source+scope baseline) to 0.032 overall. However, held-out false promotion remained 0.111, and the method proved highly conservative with a 0.692 review burden and 0.448 direct recall. A final human adjudication of 133 high-impact external coding-agent candidates was even more severe: none were safe for automatic promotion. All 11 verification-gate positives were rejected as boilerplate, shared-tool artifacts, file dumps, or non-reusable debugging traces. The authors conclude that GovMem serves primarily as a diagnostic governance design point, not a general automatic memory writer, and emphasize that agent memory write paths should be evaluated as risk-controlled evidence-governance systems.