RegimeCouncil paper proposes online routing between local and global memory feedback in LLM agents
An arXiv paper frames stale long-term memory as a governance problem and introduces a router that switches between two feedback-sharing policies without retraining the model.
In plain words
AI agents that remember past user details can act on facts that are no longer true, and they often only learn this after the damage is done. The paper asks how far that bad experience should count against memory: only for the same fact, or for all stored facts. RegimeCouncil is a small decision module that keeps both options and picks between them while the agent runs.
What was announced
The authors, writing on arXiv, formulate this online scope choice under hidden validity and delayed feedback as memory conflict governance and argue that neither fixed policy is sufficient alone. Their system keeps two advisors: a Local Advisor tracks the latest outcome for the queried fact, while a Global Advisor pools recent outcomes across all stored facts. A Meta Governor defaults to the local view to avoid cross-fact interference and switches to the global view only when pooled evidence shows a strictly positive recent utility margin, which the paper says guarantees finite-time switching under delayed feedback. No retraining of the underlying LLM is required.
Reported results come from LongMemEval update streams plus historical drift on TempLAMA and TAQA, using Qwen3-30B-A3B and GPT-4o. The authors state RegimeCouncil beats the better fixed-scope policy by 5.2-7.3 points on primary disagreement episodes and across all 30 held-out settings, and that removing any of its three modules cuts accuracy by 2-7 points. These are the authors' own measurements on their selected benchmarks, not independent replications.
What it does not show
- Evidence is limited to the paper's abstract; the full method, proofs and experimental setup were not reviewed.
- All performance figures are author-reported on the chosen benchmarks and models, with no independent confirmation.
Sources
Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 19 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.