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as the world changes, and an agent often learns that using memory hurt only through delayed feedback after acting on an outdated record.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
Figure 1: Three regimes of influence scale , i.e., the propagation scope of delayed feedback: (1) a single fact when drift is isolated; (2) a store-wide use-threshold after a cluster of corrections; and (3) a cold-start prior from pooled outcomes when the queried fact has no local history.
Figure 2: Overview of RegimeCouncil. (1) At step t , a query arrives for key kt ; whether the stored record is valid is hidden from the prompt, and the router chooses between injecting memory ( at=1 ) and using the no-memory anchor ( at=0 ). (2) The Global Advisor pools recent paired outcomes across all keys, whereas the Local Advisor reads only the latest revealed outcome on key kt . (3) The Meta Governor compares what each advisor would have scored over recent revealed episodes, follows Local by default, and switches to Global only when At>0 . (4) After a delay of at most τ steps, the episode’s paired outcome arrives and updates both advisors and the governor.
Disagreement slice ( gt=ℓt )
Full stream ( 352 ep.)
Schedule
Model
ndis
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 [0.0,3.4]
GPT-4o
134
84.3
79.1
67.2
+5.2
73.3
71.3
+2.0 [0.0,3.4]
B
Qwen
151
83.4
78.1
76.2
+5.3
71.6
69.3
+2.3 [−0.3,4.0]
GPT-4o
146
82.9
77.4
76.0
+5.5
73.0
70.7
+2.3 [−0.3,3.7]
Table 1: Main results on the two primary held-out schedules: disagreement-slice accuracy (%) and full-stream accuracy (%) with key-cluster bootstrap 95% CIs ( 5,000 replicates). RC denotes RegimeCouncil; G and L denote the Global and Local Advisors used alone; Best G/L is max(G,L) ; Gap is RC minus Best G/L (points); Agree-only removes the Meta Governor (§ 4.3 ). Jointly resampling all four evaluations yields a full-stream gap of +1.9[0.9,3.1] and a slice gap of +5.1[2.3,8.1] .
Schedule family
Model
Schedule
Gap
95% CI
block_switch
Qwen
A
+6.0
[0.6,9.1]
B
+8.4
[4.1,10.0]
GPT-4o
A
+8.5
[3.0,11.5]
B
+9.0
[3.7,10.6]
independent_switch
Qwen
A
+8.6
[5.1,10.1]
B
+9.2
[5.7,10.7]
Table 2: Disagreement-slice gap on the 54 -key pool with per-key snapshot lag matched to the 16 -key protocol ( refresh=324 ). All eight key-bootstrap 95% intervals ( 5,000 replicates) lie strictly above zero. Gap is RC minus Best G/L (points); windows stay frozen.
Variant
Qwen
GPT-4o
without Global Advisor
−2.3
−2.0
without Local Advisor
−3.7
−6.0
without Meta Governor
−6.8
−6.5
Table 3: Module ablations on Schedule A (change in full-stream accuracy, points), keeping all windows and thresholds frozen. Full accuracy levels and harmful-exposure counts appear in Table 6 .
Source
Model
Schedule
ndis
RC
Best G/L
Gap
95% CI
TempLAMA
Qwen
A
1105
85.4
78.2
+7.2
[5.8,8.9]
B
1109
86.2
78.8
+7.4
[6.0,9.1]
GPT-4o
A
786
84.2
66.2
+18.1
[15.2,20.5]
B
793
83.6
66.1
+17.5
[14.6,19.8]
TAQA
Qwen
A
856
83.4
74.6
+8.8
[6.4,10.8]
B
848
84.0
75.0
+9.0
[7.0,11.1]
Table 4: Real-world historical drift on TempLAMA ( 120 keys) and TAQA ( 100 keys), disagreement-slice accuracy (%). Switch times follow true historical change years with matched per-key snapshot lag and frozen windows. All eight key-bootstrap 95% CIs ( 5,000 replicates) lie strictly above zero.
Appendix figures & tables17 assets
Supplementary material from the paper’s appendix.
Appendix
Schedule
Model
Pre n
Pre acc.
Post n
Post ndis
Post gap
A
Qwen
96
37.5
256
110
+7.3
GPT-4o
96
43.8
256
134
+5.2
B
Qwen
96
37.5
256
151
+5.3
GPT-4o
96
43.8
256
146
+5.5
Appendix
Table 5: Revelation slice on the cached main streams (accuracy, %). Pre is the 96 episodes before any snapshot commit: Global, Local, and RegimeCouncil tie, and ndis=0 . Post gap is RC−max(G,L) on disagreements after the first commit, equal to Table 1 .
Schedule
Model
Variant
Accuracy
Harm exposure
Δ acc
A
Qwen
Full RegimeCouncil
70.2
9
—
without Global Advisor
67.9
8
−2.3
without Local Advisor
66.5
11
−3.7
without Meta Governor
63.4
8
−6.8
GPT-4o
Full RegimeCouncil
73.3
9
—
without Global Advisor
71.3
6
−2.0
Appendix
Table 6: Module ablation on Schedules A and B (full-stream accuracy, %). Each row removes one component of the full router. Without Global or without Local, the remaining advisor is used alone. Without the Meta Governor (§ 3.5 ) the row falls below either advisor alone in all four settings (on Schedule A, 6.8 points on Qwen and 6.5 on GPT-4o; 21.8 and 17.2 disagreement-slice points). Removing Local roughly doubles harmful exposure in all four settings. Δ acc is computed from unrounded accuracies.
Qwen, Schedule A
GPT-4o, Schedule A
Wm
gate 0
gate 1
gate 2
gate 0
gate 1
gate 2
4
+8.2
+6.4
+1.8
+6.0
+3.7
+3.0
8 (frozen)
+7.3
+5.5
+5.5
+5.2
+4.5
+3.7
16
+7.3
+6.4
+5.5
+4.5
+4.5
+3.0
Appendix
Table 7: Report-only disagreement gaps (points) on Schedule A. Larger gates shrink the gain. Not used for selection.
Wg
Schedule A, Qwen
Schedule A, GPT-4o
Schedule B, Qwen
Schedule B, GPT-4o
16
+9.9
+4.6
+1.8
+2.0
32 (frozen)
+7.3
+5.2
+5.3
+5.5
64
+5.7
+4.9
+3.1
+4.4
Appendix
Table 8: Report-only disagreement gaps (points) versus the best advisor at the same Wg . All twelve entries are positive; the smallest is Schedule B, Qwen, at Wg=16 ( +1.8 ). Not used for selection.
Schedule family
Schedule A, Qwen
Schedule A, GPT-4o
Schedule B, Qwen
Schedule B, GPT-4o
independent_switch
+6.7
+7.2
+6.9
+7.1
poisson_flip
+0.6
+3.1
+4.6
+3.3
Appendix
Table 9: Disagreement-slice gap (points) on two held-out schedule families, 16 -key caches, frozen windows. independent_switch is positive on all four settings; poisson_flip is positive on all four with smaller gaps.
Governor
Median gap
Worst gap
Settings <0
Switch rate
Excess harm
Wm=1
+3.97
+0.62
0/30
0.093
0.13
Wm=2
+5.73
+1.23
0/30
0.071
0.13
Wm=4
+7.09
+1.85
0/30
0.049
0.27
Wm=8 (frozen)
+7.27
+0.62
0/30
0.035
0.90
Wm=16
+7.87
+2.47
0/30
0.027
1.27
Wm=32
+7.20
+1.85
0/30
0.024
2.33
Appendix
Table 10: Every surplus-based governor outperforms the better fixed scale on all 30 held-out settings; the choice of governor mainly changes harmful exposure. Gaps are disagreement-slice points over max(G,L) (upper-middle element of the 30 sorted values; sample two-sided medians for Follow-recent-winner, CUSUM, and Wm=8 are +8.0 , +7.8 , and +7.3 in Table 17 ), alongside the worst setting. Switch rate is the fraction of consecutive disagreement episodes on which the governor changes advisor. Excess harm is the mean number of episodes per setting, beyond the safest fixed scale, in which the policy acts on memory that is worse than the anchor.
Development family
RC − follow
Harm, RC
Harm, follow
v4_default
+0.62
49
65
anticorrelated
+0.47
38
46
long_stable_flip
+0.07
94
105
sparse_need_global
+0.53
17
28
block_switch
+0.84
69
79
Appendix
Table 11: Development comparison, disagreement-slice points and harmful-episode counts, pooled over five tags and both models (refresh 96 ). The frozen rule has higher accuracy and fewer harmful episodes in every family, including the family Schedule A is drawn from. Per model the accuracy margins are +0.62 (Qwen) and +0.40 (GPT-4o).
Model
Schedule
ndis
RC
Best G/L
Gap
95% CI
Qwen
A
1105
85.4
78.2
+7.2
[5.8,8.9]
B
1109
86.2
78.8
+7.4
[6.0,9.1]
GPT-4o
A
786
84.2
66.2
+18.1
[15.2,20.5]
B
793
83.6
66.1
+17.5
[14.6,19.8]
Appendix
Table 12: TempLAMA real-drift stream ( 120 keys, real switch times, matched snapshot lag), disagreement-slice accuracy (%). Gap is RC minus Best G/L (points). The large GPT-4o gap is driven by the fixed Global and Local scales collapsing to ≈66% on the slice, not by an unusual RC level. Full-stream gaps are 3.0 and 3.1 points on Qwen and 5.4 and 5.3 points on GPT-4o (Schedules A and B). The difference from follow-recent-winner and CUSUM is within ≈0.7 points. All four key-bootstrap 95% CIs ( 5000 replicates) are above zero ( P(gap>0)=1.00 ); each ablation reduces accuracy (without Local −5.7 to −7.0 , without Global −3.0 to −5.4 , without Meta Governor −3.9 to −5.8 ). Holding refresh=96 while K grows shortens the per-key lag below the protocol and is not comparable.
Model
Schedule
ndis
RC
Best G/L
Gap
95% CI
Qwen
A
856
83.4
74.6
+8.8
[6.4,10.8]
B
848
84.0
75.0
+9.0
[7.0,11.1]
GPT-4o
A
702
81.5
69.8
+11.7
[9.1,14.7]
B
682
82.4
70.1
+12.3
[9.5,15.6]
Appendix
Table 13: TAQA real-drift stream ( 100 keys, real switch times, matched snapshot lag), disagreement-slice accuracy (%). Gap is RC minus Best G/L (points). Full-stream gaps are 3.4 and 3.5 points on Qwen and 3.7 and 3.8 points on GPT-4o (Schedules A and B). The difference from follow-recent-winner and CUSUM is within ≈0.6 points. All four key-bootstrap 95% CIs ( 5000 replicates) are above zero ( P(gap>0)=1.00 ); each ablation reduces accuracy (without Local −3.9 to −5.3 , without Global −3.4 to −3.8 , without Meta Governor −3.5 to −5.2 ).
Model
Method
Accuracy
Harm exposure
Drawdown
Qwen
No memory
48.9
0
0
Always memory
76.7
23
6
Best global ( W=16 )
76.7
16
2
Best local ( W=1 )
75.9
8
1
RegimeCouncil
78.4
8
1
GPT-4o
No memory
48.6
0
0
Appendix
Table 14: Complementary Pareto summary on the harm-exposure schedule.
Figure 3: Accuracy vs. harmful-memory exposure on the harm-exposure schedule. Blue circles: Global window. Green squares: Local (per-key). Orange diamonds: no / always memory. Red star: RegimeCouncil. Black outlines mark the best Global and best Local settings from Table 14 .
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
Appendix
Table 15: Observable-conflict selection on MemoryAgentBench 64K (held-out %, n=68 ). Not a RegimeCouncil score: both writes are in the prompt, and the choice is between generic retrieval and last-write-wins. Lifts of the selected policy over Gen@5: 35.3 points on GPT-4o and 25.0 on Qwen. Generic k=20 never certifies.
Episode (hidden gold)
Global
Local
Meta
RC
Volleyball record ( 3-2 )
use
skip
override
use (correct)
Spare screwdriver (No)
skip
use
keep Local
use (correct)
Appendix
Table 16: Two disagreement steps on Schedule A (Qwen). Global pools the last 32 paired advantages; Local reads only the current key. Meta overrides Local only if recent counterfactual surplus is strictly positive.
Case 1: 30 Held-Out
Case 2: Clustered
Case 1+2
Method
Full
Gap
Harm
sync_multi
sync_switch
Two-Stage
Follow-recent-winner
71.7
+8.0
746 (2.21%)
94.2
93.5 (37)
69.5 (65)
Fixed Share ( η=2,α=0.3 )
71.7
+8.0
746 (2.21%)
94.2
93.5 (37)
66.8 (74)
CUSUM on surplus
71.3
+7.8
746 (2.21%)
93.1
92.8 (41)
67.1 (70)
Local + tie-deferral (no Meta)
72.4
+9.7
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)
Appendix
Table 17: Comparison across Case 1 ( 30 held-out settings, 33,792 episodes), Case 2 ( 54 -key clustered shifts sync_multi and sync_switch ), and Case 1+2 ( 54 -key two-stage stream with asynchronous flips followed by store-wide flips; disagreement-slice accuracy in % and total harmful episodes). Default RegimeCouncil achieves the lowest harmful exposure on the 30 settings ( 707 ) and outperforms standalone tie-deferral by +27.9 points under clustered corrections while reducing harmful episodes by 59% on sync_switch ( 37 vs. 91 ) and 26% on two-stage streams ( 75 vs. 102 ); RC-refined achieves the highest accuracy across both regimes.
Mean Full-Stream Accuracy (%)
Gain vs. Always
Snapshot Commit Interval
Always
max(G,L)
RC (def.)
RC-ref.
RC (def.)
RC-ref.
1 or 16 ep. ( 1 pass, write-back)
93.0
92.6
91.6
92.7
−1.4
−0.3
32 episodes ( 2 passes)
87.7
86.4
86.9
87.8
−0.8
+0.1
64 episodes ( 4 passes)
78.9
77.5
79.4
80.3
+0.5
+1.4
96 episodes ( 6 passes, main)
70.1
69.6
71.9
72.8
+1.8
+2.7
192 episodes ( 12 passes)
51.2
56.7
56.7
56.8
+5.5
+5.6
Appendix
Table 18: Varying the snapshot commit interval at fixed outcome delay τ=1 across Schedules A, B, and C on both models (mean full-stream accuracy, %). Since each of the 16 keys is visited once per 16 -episode pass, commit intervals ≤16 correspond to immediate write-back before the key’s next visit.
Key Pool (Commit =6K )
Router Variant
τ=1
τ=4
τ=8
τ=16
τ=48
τ=96
16-key pool (commit 96)
RegimeCouncil (default)
+5.8
+3.6
−1.2
−4.9
−11.1
−24.4
RegimeCouncil ( RC-refined )
+8.3
+7.7
+6.9
+5.9
+10.4
+6.2
54-key pool (commit 324)
RegimeCouncil (default)
+8.6
+7.7
+6.8
+4.8
+0.6
−7.3
RegimeCouncil ( RC-refined )
+9.9
+9.6
+9.4
+9.4
+8.6
+7.5
TempLAMA 120 / TAQA 100
RegimeCouncil (default)
+11.5
+11.0
+10.1
+8.8
+2.6
0.0
RegimeCouncil ( RC-refined )
+12.3
+12.2
+11.8
+11.2
+9.4
+8.4
Appendix
Table 19: Varying the outcome revelation delay τ at fixed commit interval 6K (disagreement-slice gap vs. max(G,L) , points). For each pool, the first row reports the default RegimeCouncil and the second row reports RC-refined (tie-deferring Local with the Meta Governor).
Evaluation Condition / Cost Metric
RegimeCouncil (default)
Baseline Comparison
In-stream random delay ( τt∼Uniform{1..8} / {1..16} )
+6.7 / +5.9 median slice gain (149/150 & 140/150 runs positive)
Outperforms Fixed Share & Follow (133/150 positive at {1..16} )
Sparse feedback (75% / 50% of episodes revealed)
+5.7 / +3.2 median slice gain ( +9.8 / +8.7 with RC-refined)
Router CPU time & throughput (Python, 33,792 decisions)
2.29 μ s/decision ( 4.4×105 decisions/s)
Global alone: 0.88 μ s/decision (+1.41 μ s router overhead)
Appendix
Table 20: Performance under realistic feedback conditions across all 30 held-out settings and per-episode computational overhead relative to the Global Advisor alone.
Wuhan University1, Wuhan, China1 · The Hong Kong University of Science and Technology3, Hong Kong, China23 · The Chinese University of Hong Kong2, Hong Kong, China23