RegimeCouncil: Paired-Evidence Routing For Memory Conflict Governance In LLM Agents
Organizations: Huawei Noah’s Ark Lab, Hong Kong · City University of Hong Kong, Hong Kong
Abstract
Language agents rely on long-term memory to personalize responses across extended interactions, yet stored facts quietly go stale as the world changes, and an agent often learns that using memory hurt only through delayed feedback after acting on an outdated record. Before the memory store is updated, how far should that negative feedback propagate: should it suppress memory use only on future queries for that same fact (a local policy), or also across other stored facts (a global policy)? Existing memory systems implicitly fix one of these two sharing scopes at design time. We formulate this online scope choice under hidden validity and delayed feedback as memory conflict governance, and prove that neither fixed policy is sufficient on its own. We then propose RegimeCouncil, a lightweight paired-evidence router that retains both scopes as advisors and arbitrates between them online without retraining the LLM. Using revealed comparisons between answering with memory and answering without memory (the no-memory anchor), a Local Advisor tracks the latest outcome on the queried fact alone, while a Global Advisor pools recent outcomes across all stored facts. A Meta Governor follows the Local Advisor by default to avoid cross-fact interference and switches to the Global Advisor only when global pooling accumulates a strictly positive recent utility margin, guaranteeing finite-time switching under delayed feedback. Across LongMemEval update streams and real-world historical drift on TempLAMA and TAQA with Qwen3-30B-A3B and GPT-4o, RegimeCouncil outperforms the better fixed-scope policy by 5.2-7.3 points on primary disagreement episodes and across all 30 held-out settings, while ablations show that removing any of its three modules reduces accuracy by 2-7 points.
Figures & tables
| Disagreement slice ( ) | Full stream ( ep.) | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| Schedule | Model | RC | Best G/L | Agree-only | Slice Gap | RC | Best G/L | Full Gap [95% CI] | |
| A | Qwen | 110 | 87.3 | 80.0 | 65.5 | +7.3 | 70.2 | 67.9 | +2.3 |
| GPT-4o | 134 | 84.3 | 79.1 | 67.2 | +5.2 | 73.3 | 71.3 | +2.0 | |
| B | Qwen | 151 | 83.4 | 78.1 | 76.2 | +5.3 | 71.6 | 69.3 | +2.3 |
| GPT-4o | 146 | 82.9 | 77.4 | 76.0 | +5.5 | 73.0 | 70.7 | +2.3 | |
| Schedule family | Model | Schedule | Gap | 95% CI |
|---|---|---|---|---|
| block_switch | Qwen | A | ||
| B | ||||
| GPT-4o | A | |||
| B | ||||
| independent_switch | Qwen | A | ||
| B |
| Variant | Qwen | GPT-4o |
|---|---|---|
| without Global Advisor | ||
| without Local Advisor | ||
| without Meta Governor |
| Source | Model | Schedule | RC | Best G/L | Gap | 95% CI | |
|---|---|---|---|---|---|---|---|
| TempLAMA | Qwen | A | 1105 | 85.4 | 78.2 | ||
| B | 1109 | 86.2 | 78.8 | ||||
| GPT-4o | A | 786 | 84.2 | 66.2 | |||
| B | 793 | 83.6 | 66.1 | ||||
| TAQA | Qwen | A | 856 | 83.4 | 74.6 | ||
| B | 848 | 84.0 | 75.0 |
Appendix figures & tables17 assets
Supplementary material from the paper’s appendix.
Appendix
| Schedule | Model | Pre | Pre acc. | Post | Post | Post gap |
|---|---|---|---|---|---|---|
| A | Qwen | 96 | 37.5 | 256 | 110 | |
| GPT-4o | 96 | 43.8 | 256 | 134 | ||
| B | Qwen | 96 | 37.5 | 256 | 151 | |
| GPT-4o | 96 | 43.8 | 256 | 146 |
| Schedule | Model | Variant | Accuracy | Harm exposure | acc |
|---|---|---|---|---|---|
| A | Qwen | Full RegimeCouncil | 70.2 | 9 | — |
| without Global Advisor | 67.9 | 8 | |||
| without Local Advisor | 66.5 | 11 | |||
| without Meta Governor | 63.4 | 8 | |||
| GPT-4o | Full RegimeCouncil | 73.3 | 9 | — | |
| without Global Advisor | 71.3 | 6 |
| Qwen, Schedule A | GPT-4o, Schedule A | |||||
|---|---|---|---|---|---|---|
| gate | gate | gate | gate | gate | gate | |
| 4 | ||||||
| 8 (frozen) | ||||||
| 16 | ||||||
| Schedule A, Qwen | Schedule A, GPT-4o | Schedule B, Qwen | Schedule B, GPT-4o | |
|---|---|---|---|---|
| 16 | ||||
| 32 (frozen) | ||||
| 64 |
| Schedule family | Schedule A, Qwen | Schedule A, GPT-4o | Schedule B, Qwen | Schedule B, GPT-4o |
|---|---|---|---|---|
| independent_switch | ||||
| poisson_flip |
| Governor | Median gap | Worst gap | Settings | Switch rate | Excess harm |
|---|---|---|---|---|---|
| (frozen) | |||||
| Development family | RC follow | Harm, RC | Harm, follow |
|---|---|---|---|
| v4_default | |||
| anticorrelated | |||
| long_stable_flip | |||
| sparse_need_global | |||
| block_switch |
| Model | Schedule | RC | Best G/L | Gap | 95% CI | |
|---|---|---|---|---|---|---|
| Qwen | A | 1105 | ||||
| B | 1109 | |||||
| GPT-4o | A | 786 | ||||
| B | 793 |
| Model | Schedule | RC | Best G/L | Gap | 95% CI | |
|---|---|---|---|---|---|---|
| Qwen | A | 856 | ||||
| B | 848 | |||||
| GPT-4o | A | 702 | ||||
| B | 682 |
| Model | Method | Accuracy | Harm exposure | Drawdown |
|---|---|---|---|---|
| Qwen | No memory | 48.9 | 0 | 0 |
| Always memory | 76.7 | 23 | 6 | |
| Best global ( ) | 76.7 | 16 | 2 | |
| Best local ( ) | 75.9 | 8 | 1 | |
| RegimeCouncil | 78.4 | 8 | 1 | |
| GPT-4o | No memory | 48.6 | 0 | 0 |
| Model | Gen@5 | Gen@20 | LWW@5 | LWW@20 | Selected | Guarded | Online Meta |
|---|---|---|---|---|---|---|---|
| GPT-4o | 19.1 | 20.6 | 55.9 | 54.4 | LWW@20 | 54.4 | 54.4 |
| Qwen | 17.6 | 17.6 | 42.6 | 39.7 | LWW@5 | 42.6 | 41.2 |
| Episode (hidden gold) | Global | Local | Meta | RC |
|---|---|---|---|---|
| Volleyball record ( ) | use | skip | override | use (correct) |
| Spare screwdriver (No) | skip | use | keep Local | use (correct) |
| Case 1: Held-Out | Case 2: Clustered | Case 1+2 | ||||
|---|---|---|---|---|---|---|
| Method | Full | Gap | Harm | sync_multi | sync_switch | Two-Stage |
| Follow-recent-winner | 71.7 | 746 (2.21%) | 94.2 | 93.5 (37) | 69.5 (65) | |
| Fixed Share ( ) | 71.7 | 746 (2.21%) | 94.2 | 93.5 (37) | 66.8 (74) | |
| CUSUM on surplus | 71.3 | 746 (2.21%) | 93.1 | 92.8 (41) | 67.1 (70) | |
| Local + tie-deferral (no Meta) | 72.4 | 752 (2.23%) | 66.3 | 91.3 (91) | 69.7 (102) | |
| RegimeCouncil (default) | 71.6 | +7.3 | 707 (2.09%) | 94.2 | 93.5 (37) | 68.2 (75) |
| Mean Full-Stream Accuracy (%) | Gain vs. Always | |||||
|---|---|---|---|---|---|---|
| Snapshot Commit Interval | Always | RC (def.) | RC-ref. | RC (def.) | RC-ref. | |
| 1 or 16 ep. ( pass, write-back) | 93.0 | 92.6 | 91.6 | 92.7 | ||
| 32 episodes ( passes) | 87.7 | 86.4 | 86.9 | 87.8 | +0.1 | |
| 64 episodes ( passes) | 78.9 | 77.5 | 79.4 | 80.3 | +0.5 | +1.4 |
| 96 episodes ( passes, main) | 70.1 | 69.6 | 71.9 | 72.8 | +1.8 | +2.7 |
| 192 episodes ( passes) | 51.2 | 56.7 | 56.7 | 56.8 | +5.5 | +5.6 |
| Key Pool (Commit ) | Router Variant | ||||||
| 16-key pool (commit 96) | RegimeCouncil (default) | ||||||
| RegimeCouncil ( RC-refined ) | +8.3 | +7.7 | +6.9 | +5.9 | +10.4 | +6.2 | |
| 54-key pool (commit 324) | RegimeCouncil (default) | ||||||
| RegimeCouncil ( RC-refined ) | +9.9 | +9.6 | +9.4 | +9.4 | +8.6 | +7.5 | |
| TempLAMA 120 / TAQA 100 | RegimeCouncil (default) | ||||||
| RegimeCouncil ( RC-refined ) | +12.3 | +12.2 | +11.8 | +11.2 | +9.4 | +8.4 |
| Evaluation Condition / Cost Metric | RegimeCouncil (default) | Baseline Comparison |
|---|---|---|
| In-stream random delay ( / ) | +6.7 / +5.9 median slice gain (149/150 & 140/150 runs positive) | Outperforms Fixed Share & Follow (133/150 positive at ) |
| Sparse feedback (75% / 50% of episodes revealed) | +5.7 / +3.2 median slice gain ( +9.8 / +8.7 with RC-refined) | Outperforms Follow & Fixed Share ( / median gain) |
| Noisy feedback (5% / 10% outcome flip probability) | +3.5 / +1.0 median slice gain | Outperforms Follow & Fixed Share ( / median gain) |
| Executed-action-only bandit (IPW ; no call) | 68.9% full-stream accuracy ( 31.3 harmful episodes/setting) | Outperforms Global ( 63.2% acc., higher in 30/30 settings, 60.6 harm), Follow ( 68.5% acc., higher in 23/30, ), & Local ( 68.5% acc., higher in 18/30) |
| User-facing LLM calls & latency | 1 call (+0 ms added latency) | Global: 1 user-facing call (+1 shared background call for ) |
| Router CPU time & throughput (Python, decisions) | 2.29 s/decision ( decisions/s) | Global alone: 0.88 s/decision (+1.41 s router overhead) |