ARTFEED — Contemporary Art Intelligence

Governed Persistent Memory: Source-Bound State Semantics for Long-Horizon Agents

ai-technology · 2026-08-15

A recent submission to arXiv (2608.12476) presents Governed Persistent Memory (GPM), an auditable bitemporal state-transition framework aimed at resolving issues related to contradictory, outdated, retracted, deleted, or stale records in long-term agent memory. This model incorporates features such as source-bound admission, a derived lifecycle state, current public barriers, and a fail-closed structured release. It includes five executable clauses that ensure ledger integrity, source binding, conflict isolation, prevention of revival post-retraction or deletion, and precise claim closure at a verified head. In tests on a hash-frozen 3,600-case GPM-ReleaseBench, GPM achieved complete outcomes, with the best of three simple policies matching 1,800/3,600 and yielding unmatched releases in 50% of violation scenarios. Additionally, a separate evaluation of a sealed end-to-end service assesses real ingestion and release across eight query families.

Key facts

  • Paper arXiv:2608.12476 introduces Governed Persistent Memory (GPM).
  • GPM is an auditable bitemporal state-transition model.
  • It includes source-bound admission, derived lifecycle state, current public barriers, and fail-closed structured release.
  • Five executable clauses cover ledger integrity, source binding, conflict isolation, non-revival, and exact claim closure.
  • GPM-ReleaseBench has 3,600 cases and is hash-frozen.
  • GPM matches all complete outcomes on GPM-ReleaseBench.
  • The strongest simple complete policy matches 1,800/3,600 and makes unmatched releases on 50% of violation cases.
  • A sealed end-to-end service evaluation exercises real ingestion and release across eight query families.

Entities

Institutions

  • arXiv

Sources