Agents inspect a memory constraint's provenance in one episode out of five, and act on withdrawn rules 77% of the time
A single-author study submitted 2026-08-26 (arXiv 2608.25553) models supersession in inherited agent memory: a constraint that was true when written and has since been withdrawn by a newer authoritative record. Given a two-record verification budget, agents checked the constraint's provenance path in about one episode in five, and produced stale-consistent decisions in 77.3%, 74.7% and 74.7% of episodes across a primary run, a fresh-wording replication and a held-out domain. Re-assigning one of the two slots to the critical provenance path raised current-record-consistent decisions by 74.0, 72.7 and 61.3 points, positive in six of six models in every run, and changed nothing when the record agreed with the memory. The fix costs no extra budget, only a different allocation.
↳ Follow the thread