Organizations: Nankai University · Tsinghua University · Huazhong University of Science and Technology · University of the Chinese Academy of Sciences · vivo AI Lab
Long-horizon dialogue agents need to keep user memories stable enough for personalization while adapting when the user genuinely changes. Existing memory systems often append every observation or overwrite old entries by heuristic rules, which makes transient noise hard to distinguish from real belief drift. We introduce PreCoMem, a memory consolidation framework inspired by predictive coding that represents memory as confidence-weighted beliefs about the user's state. Its core signal, Effective Surprisal, combines semantic deviation and contradiction evidence, then downweights this signal when retrieval is diffuse and unreliable. According to effective surprisal, a two-threshold gate maps each turn to one of three deterministic updates: MAINTAIN reinforces supported beliefs, PROFILE stores ambiguous cues as low-confidence hypotheses, and CORRECT softly decays reliably contradicted facts. The PROFILE stage acts as an explicit waiting room, preventing premature commitment to noisy evidence. Experiments on LoCoMo, LongMemEval-S, and PersonaMem-v2 show consistent state-of-the-art accuracy; our new DiMoBench confirms that PreCoMem adapts to non-stationary belief drift without over-reacting to noise. The code is available at https://github.com/ZhongWishing/PreCoMem.
Figures & tables
Figure 1: Memory evolution case study (coffee → tea). PreCoMem autonomously manages the lifecycle of facts and hypotheses, achieving a stable transition between conflicting belief states without manual intervention.
Figure 2: The PreCoMem framework. Two active phases operate over a shared memory store. Phase 1, Memory Leveraging : a confidence-modulated retrieval score combines semantic, lexical, and topic signals to select Top- K beliefs Ct for response generation. Phase 2, Memory Evolution : the effective surprisal Seff and a two-threshold gate route each turn to MAINTAIN , PROFILE , or CORRECT , illustrated by representative low/medium/high- Seff turns. Memory Store : an editable Belief Store B of FACT / HYPO nodes with confidence weights, linked by provenance to an immutable Episodic Cache E under a shared topic index; belief items are updated independently, and the drawn links are provenance pointers and topic groupings (Appendix Section C.11 ).
LoCoMo (1,540 questions; four splits)
Method
Avg. Tokens
Single-Hop
Multi-Hop
Temporal
Open-Domain
Overall
Backbone: gpt-4o-mini
MemoryOS
5.5k
66.68 ± 0.85
54.83 ± 0.87
52.90 ± 0.59
32.16 ± 0.11
59.58 ± 0.06
Mem0
1.6k
62.17 ± 0.77
52.91 ± 0.70
59.44 ± 0.37
33.89 ± 0.96
58.36 ± 0.31
A-Mem
2.7k
45.15 ± 0.53
23.99 ± 0.46
53.67 ± 0.23
30.71 ± 0.84
42.18 ± 0.65
GAM
3.5k
68.24 ± 0.76
55.47 ± 0.43
58.19 ± 0.54
33.50 ± 0.09
61.62 ± 0.22
Table 1: Main evaluation on LoCoMo and LongMemEval - S. Accuracies are reported as mean ± std over three seeds; within each backbone block, bold marks the best score per column and underline marks the second-best. Avg. Tokens reports the average response-model input token usage per query. LongMemEval - S columns are question types: SSU/SSA/SSP = single-session user/assistant/preference, MS = multi-session, KU = knowledge update, and TR = temporal reasoning.
Figure 3: Backbone generalization and PersonaMem - v2 system comparison. (a)–(d) Scale-vs-mechanism comparison on LoCoMo, LongMemEval - S, PersonaMem - v2 (32k), and PersonaMem - v2 (128k). Within each group the answer model is fixed (Qwen3.5-35B, Qwen3.5-122B, or GPT-5-Chat): the left bar runs it with its native O-Mem memory, while the right bar keeps the same answer model and delegates memory construction to PreCoMem on a one-tier-smaller backbone (Qwen3.5-9B, Qwen3.5-35B, GPT-5-mini); the protocol is detailed in Appendix Section C.2 . (e)–(f) Head-to-head with the five strongest baselines on PersonaMem - v2 under gpt-4.1 , in both MCQ and Open-Ended formats. Error bars are ± std over three seeds.
Figure 4: Component ablation and DiMoBench dynamics. (a) Component ablation on LoCoMo and LongMemEval - S. (b) Moving-average accuracy under implicit drift on DiMoBench. (c) Operator-level ablation on DiMoBench.
Figure 5: Hyperparameter and memory-scalability analysis. (a) LoCoMo accuracy–efficiency frontier as the retrieval budget k varies; overall accuracy peaks at k=10 , beyond which additional context yields diminishing returns and dilutes retrieval reliability ρt . (b, c) Memory scalability under long-horizon belief drift: raw memory grows linearly, while effective memory quickly plateaus on both DiMoBench and DiMoBench - XL (panel (c) shows one representative persona).
Method
Per-item conf.
Nrawend
Neffend
Top-10 avg. w
LoCoMo Overall
O-Mem [ 40 ]
✗ ( Add / Ignore / Update )
232
–
–
70.88
EverMemOS [ 10 ]
✗ ( [ts,te] expiry)
264
–
–
70.13
PreCoMem (A0, default)
✓ ( wt∈[0,1] )
287
95.7
0.81
73.57
PreCoMem (A1, w<0.05 prune)
✓ ( wt∈[0,1] )
168
95.6
0.83
73.51
Table 2: Memory footprint vs. accuracy on the ten LoCoMo dialogues (mean turns Nˉ=588 ). All numbers are measured under the same backbone ( gpt-4o-mini ) and embedding model. “Per-item conf.” indicates whether the system maintains a continuous confidence weight per stored item; under discrete commit decisions, the effective count Neff and the average Top-10 weight are not defined. PreCoMem (A1) is the optional pruning configuration that drops items with wt<0.05 .
Appendix figures & tables30 assets
Supplementary material from the paper’s appendix.
Appendix
System
Overall ↑
Memory ↑
No memory (gpt-4.1-mini) †
0.368
0.203
A-Mem †
0.378
0.220
Nemori †
0.385
0.231
Mem0-Graph †
0.424
0.284
AWE †
0.430
0.296
PreCoMem
0.458 ± 0.006
0.329 ± 0.009
Appendix
Table 3: AMemGym (official v1.base , 20 profiles × 11 periods, gpt-4.1-mini for all memory systems). Overall accuracy and normalized Memory score; PreCoMem is mean ± std over three seeds. Rows marked † are the values reported in the AMemGym paper under the same backbone; AWE was additionally run locally only for the probe metrics quoted in the text.
Variant
Overall ↑
Memory ↑
Write fail. ↓
Read fail. ↓
Util. fail. ↓
Full PreCoMem
0.458
0.329
.300
.170
.072
w/o PROFILE
0.444
0.309
.320
.165
.071
w/o CORRECT
0.439
0.303
.312
.178
.071
w/o MAINTAIN
0.418
0.273
.310
.190
.082
Appendix
Table 4: Operator ablation on AMemGym (mean over three seeds). Top: the benchmark’s native scores and failure rates (the three failure rates sum to 1− Overall). Bottom: the state-query probes defined in the text (203 changed-state and 397 stable-state probes per seed). Paired 95% CIs of Full minus each variant exclude zero on every metric.
Table 5: LoCoMo - Plus by constraint type (constraint-consistency accuracy, %, GPT-4o for all systems; mean ± std over three seeds; Overall is the question-weighted aggregate over the 401 questions). Baseline overall scores are those reported in the LoCoMo - Plus paper.
LoCoMo
LongMemEval - S
PersonaMem - v2
Method
gpt-4o-mini
gpt-4.1
gpt-4o-mini
gpt-4.1
32k
128k
MemoryOS
5.5
5.2
7.2
6.8
5.8
7.2
Mem0
1.6
1.8
1.8
2.0
1.9
2.1
A-Mem
2.7
2.9
3.3
3.6
3.2
3.8
GAM
3.5
3.4
4.3
4.2
3.7
4.4
MemBox
3.3
3.4
3.9
4.0
3.6
4.3
Appendix
Table 6: Response-model input tokens per query (thousands) across benchmarks and backbones; PersonaMem - v2 under gpt-4.1 , averaged over MCQ and Open-Ended queries. Token usage is reported for all nine baselines, including those whose accuracy is not tabulated on a given benchmark in Table 1 or Figure 3 . The cross-seed standard deviation is below 0.15 k for every entry.
Method
Avg. Tokens
SSU
SSA
SSP
MS
KU
TR
Overall
MemoryOS
7.2k
58.21 ± 0.62
59.60 ± 0.92
46.03 ± 0.33
53.68 ± 0.35
54.71 ± 0.23
51.45 ± 0.06
54.85 ± 0.80
MemOS
2.7k
70.15 ± 0.57
68.95 ± 0.67
56.02 ± 0.96
62.53 ± 0.17
66.54 ± 0.26
64.70 ± 0.41
64.66 ± 0.78
TiMem
3.6k
81.79 ± 0.93
75.01 ± 0.36
69.99 ± 0.30
72.97 ± 0.91
77.63 ± 0.46
85.27 ± 0.08
75.51 ± 0.90
GAM
4.3k
77.61 ± 0.94
79.52 ± 0.72
68.18 ± 0.31
76.16 ± 0.43
74.09 ± 0.65
76.50 ± 0.48
76.19 ± 0.83
MemBox
3.9k
79.40 ± 0.67
83.26 ± 0.77
71.19 ± 0.46
76.16 ± 0.34
75.44 ± 0.53
78.55 ± 0.11
77.37 ± 0.39
O-Mem
3.7k
88.97 ± 0.99
82.28 ± 0.42
70.55 ± 0.45
79.61 ± 0.01
84.68 ± 0.12
78.08 ± 0.69
81.78 ± 0.36
Appendix
Table 7: LongMemEval - S under gpt-4o-mini . Mean ± std over three seeds; bold best, underline second-best. Avg. Tokens is the average response-model input token usage per query, as in Table 1 . Columns are question types: SSU/SSA/SSP = single-session user/assistant/preference, MS = multi-session, KU = knowledge update, and TR = temporal reasoning.
Dataset
Question Type
#Q
LLM-judge (%)
Human (%)
Agreement (%)
Cohen’s κ
LoCoMo
Single-Hop
109
81.41
82.73
98.5
0.949
Multi-Hop
30
68.49
70.40
98.0
0.953
Temporal
40
66.21
67.65
98.2
0.959
Open-Domain
20
39.63
41.55
97.5
0.948
Overall
199
71.96
73.30
98.3
0.957
LongMemEval - S
SSU
70
91.58
92.86
97.4
0.819
Appendix
Table 8: Per-category LLM-judge validation on LoCoMo, LongMemEval - S, and DiMoBench. For each question type we report the LLM judge’s verdict on PreCoMem outputs ( gpt-4o-mini backbone), the expert human re-annotation accuracy on the same outputs, the agreement rate, and Cohen’s κ , the chance-corrected agreement between the LLM judge and the human rater. For LongMemEval - S, SSU/SSA/SSP denote single-session user/assistant/preference, MS denotes multi-session, KU denotes knowledge update, and TR denotes temporal reasoning. Following the human-agreement check of Wu et al. [43] , we find the gpt-4.1-mini primary judge highly consistent with human raters across all three benchmarks; an additional cross-method spot-check on EverMemOS and TiMem outputs (described in the text) confirms the protocol is not calibrated to PreCoMem outputs alone, so the LLM-judged numbers in Table 1 faithfully reflect human assessments.
Config
Created
Promoted
Expired
Promotion rate
Promotion precision
PreCoMem (full)
123
43
81
0.34
0.78
w/o C1
149
42
108
0.28
0.62
w/o C2
138
40
93
0.30
0.70
w/o C3
0
0
0
0.00
–
Appendix
Table 9: Hypothesis lifecycle statistics on LoCoMo (averaged over 10 samples).
Figure 6: Hypothesis lifecycle under PreCoMem. A hypothesis created by PROFILE (Stage 1) enters multi-turn verification (Stage 2): consistent support raises its confidence toward the promotion threshold τpromote (Stage 3 → Stage 4, becoming a stable core-profile belief that ranks above hypotheses in confidence-weighted retrieval), while no/weak support lets its weight decay; once wt falls below the prune threshold the hypothesis is removed from the active pool (Stage 5a); a committed FACT that is later reliably contradicted is instead softly decayed and kept for provenance (Stage 5b). The horizontal axis is confidence, from low confidence through the promotion threshold to stable high confidence.
Variant
LoCoMo
LongMemEval - S
Full PreCoMem
73.57
84.28
w/o dsem ( λ=0 )
67.83 ( −5.74 )
77.65 ( −6.63 )
w/o dconf ( λ=1 )
63.49 ( −10.08 )
72.71 ( −11.57 )
Appendix
Table 10: Fine-grained ablation of the raw-surprisal components. Accuracy (%) on LoCoMo and LongMemEval - S ( gpt-4o-mini backbone). w/o dsem ( λ=0 ) keeps only the NLI-based contradiction signal; w/o dconf ( λ=1 ) keeps only the semantic-gap signal. Both signals matter, but dconf contributes nearly twice as much because it directly detects reliable negation.
Verifconf
Sample-1 Overall
Δ
cross-encoder/nli-deberta-v3-base (default)
76.19±0.42
–
facebook/bart-large-mnli
75.74±0.38
−0.45
roberta-large-mnli
75.40±0.51
−0.79
random πt∼Beta(2,5) (control)
65.08±0.79
−11.11
Appendix
Table 11: NLI verifier sanity check on LoCoMo Sample 1. Sample-1 LLM-judge accuracy (%) when Verifconf is replaced. The two alternative MNLI-trained verifiers stay within 1 point of the default; a randomized control collapses by 11 points, confirming that dconf is load-bearing yet that PreCoMem is not coupled to a specific verifier checkpoint.
Variant
LoCoMo
LongMemEval - S
Full PreCoMem
73.57 ± 0.28
84.28 ± 0.05
Single-threshold gate (no PROFILE band; medium band → CORRECT )
69.92 ± 0.29
81.16 ± 0.12
Frequency-matched random routing (M/P/C =56.2/27.4/16.4% )
61.83 ± 1.25
72.94 ± 1.08
Read-time weighting off, lifecycle confidence on
71.80 ± 0.43
83.56 ± 0.23
Lifecycle confidence off, read-time weighting on
52.88 ± 0.69
61.65 ± 0.68
Both off ( = w/o CW)
43.42 ± 0.94
49.65 ± 0.91
Appendix
Table 12: Route-matched and factorial controls (LLM-judge accuracy, gpt-4o-mini , mean ± std over three seeds).
LoCoMo
LongMemEval - S
PreCoMem (bounded scalars + closed-form updates)
85.06 ± 0.07
89.39 ± 0.76
LLM direct write
82.31 ± 1.42
86.24 ± 1.63
Paired difference
+2.75
+3.15
Appendix
Table 13: LLM-direct-write baseline (Overall LLM-judge accuracy, gpt-4.1 , mean ± std over three seeds; the paired difference is the mean over seeds of PreCoMem minus direct write).
Figure 7: Top- K sensitivity on LoCoMo ( gpt-4.1 backbone). Per-split accuracy as k varies in {1,…,20} . The dashed vertical line marks our default k=10 .
k
Single-Hop
Multi-Hop
Temporal
Open-Domain
Overall
1
80.5
71.2
72.4
39.6
75.2
5
88.7
78.6
80.2
45.5
82.3
10
90.41
83.48
84.74
48.23
85.06
15
89.7
82.0
82.7
46.9
84.3
20
87.9
80.6
80.3
45.8
82.5
Appendix
Table 14: Accuracy (%) at different retrieval sizes k on LoCoMo ( gpt-4.1 backbone). Best results are in bold .
Figure 8: Gating threshold sensitivity. LoCoMo accuracy (Sample 1) over a 5×5 grid of (θlow,θhigh) with step 0.05 (25 configurations in total). The red box marks the deployed default (0.20,0.60) , which sits at the center of a high-performing plateau; the four immediate neighbors of the peak stay within 1.3 – 2.6 accuracy points, while accuracy degrades smoothly toward the grid corners.
Backbone
Method
Single
Multi
Temp.
Open
Avg.
gpt-4o-mini
PreCoMem (fixed)
81.41
68.49
66.21
39.63
73.57
PreCoMem (DT)
82.27
69.31
67.04
43.55
74.55
gpt-4.1
PreCoMem (fixed)
90.41
83.48
84.74
48.23
85.06
PreCoMem (DT)
91.42
85.16
85.81
47.78
86.21
Appendix
Table 15: Adaptive vs. fixed thresholds on LoCoMo (accuracy %, gpt-4o-mini and gpt-4.1 ). PreCoMem (DT) personalizes (θlow,θhigh) via the user stability index in Equations 36 and 37 ; PreCoMem (fixed) uses (0.2,0.6) . DT improves the overall average on both backbones, with the largest gain on the drift-heavy Open category under gpt-4o-mini , without any retraining.
Figure 9: DiMoBench six-stage construction pipeline. A generator (GPT-4o) produces personas, life skeletons, dialogues, and QA pairs; an independent verifier (Gemini-2.5-Pro) cross-validates each QA pair along six criteria, and a 20% stratified human audit with κ≥0.75 provides final quality control.
Stage
Name
Model
Output
1
Persona construction
GPT-4o
User profile + core trait slot
2
Life skeleton & turn
GPT-4o
Event timeline + drift trigger
3
Dialogue generation
GPT-4o
35 sessions × 10 turns
4
QA generation
GPT-4o
200 QA pairs across 4 categories
5
Consistency verification
Gemini-2.5-Pro
Cross-validated QA (6 criteria)
6
Human–machine review
Human + LLM
20% stratified audit, κ≥0.75
Appendix
Table 16: DiMoBench construction pipeline. Generator and verifier are deliberately different models; a 20% stratified human review provides final quality control.
t
Nraw
Neff
Nactive
Nstable
Nweak
Ndepr
Neff/Nraw
100
26
21.0
13
8
3
2
0.81
200
44
25.8
16
11
7
7
0.59
280†
57
26.5
17
11
9
20
0.46
300
60
26.7
17
12
9
22
0.45
400
73
27.4
17
12
8
36
0.38
500
81
28.5
18
13
8
42
0.35
Appendix
Table 17: DiMoBench - XL memory trajectory (representative persona). Every 100 turns (and at the three drift points) on one of the five DiMoBench - XL personas; per-persona deviations from this trajectory stay within ±1.4 on Neff throughout the run. The four count columns partition Bt by confidence: Nactive ( wt≥0.8 ), Nstable ( 0.5≤wt<0.8 ), Nweak ( 0.1≤wt<0.5 ) and Ndepr ( wt<0.1 ); the last column is the ratio Neff/Nraw . Drift turns ( t∈{280,580,870} ) are marked with † .
Turn window
#QA
Acc. (%)
Avg Top-10 w
Contam.@10 (%)
050 – 250
80
84.3
0.82
0.4
250 – 500
100
83.6
0.80
0.6
500 – 750
100
83.1
0.79
0.9
750 – 1000
100
82.4
0.78
1.2
1000 – 1050
40
82.0
0.78
1.3
Appendix
Table 18: QA accuracy and retrieval quality on DiMoBench - XL across time windows. Numbers are averaged across the five DiMoBench - XL personas; per-window standard deviations across personas are ≤1.0 accuracy point and ≤0.02 for the Top-10 weight, and are omitted for readability. Average Top-10 retrieval confidence stays at or above 0.78 throughout 1050 turns; Overall accuracy decays by only 2.3 points.
Setting
Nrawend
Neffend
Contam.@10 (%)
QA Δ (%)
DiMoBench (350 t.)
62
27.6
0.6
− 0.6
DiMoBench - XL (1050 t.)
131
28.2
1.0
− 2.3
Appendix
Table 19: Memory scalability headline metrics. Nraw grows linearly with the conversation, but Neff converges to a stable plateau and the contamination of deprecated FACTs in Top-10 retrieval stays at or below 1.3% in every window even after 1050 turns and three drifts. Both rows average over their five personas; per-persona deviations are ≤0.5 on Neffend and ≤0.4 on Contam.@10.
Dataset
k
Score-mass (%)
Top- k depr. (%)
Avg retrieved w
DiMoBench
5
0.21
0.0
0.84
DiMoBench
10
0.34
0.6
0.81
DiMoBench
25
0.62
1.7
0.74
DiMoBench - XL
5
0.36
0.2
0.81
DiMoBench - XL
10
0.58
1.0
0.78
DiMoBench - XL
25
0.94
2.6
0.71
Appendix
Table 20: Retrieval contamination by deprecated FACTs ( wt<0.1 ). Even at k=25 , deprecated items occupy at most 2.6% of Top- k slots and contribute under 1% of the score mass.
Persona
Acc. (%)
Nrawend
Neffend
Contam.@10 (%)
P1 (canonical)
84.1
62
27.6
0.60
P2
83.2
65
27.8
0.80
P3
84.8
59
27.3
0.40
P4
84.4
61
27.5
0.55
P5
83.6
63
27.8
0.65
Mean
84.0
62
27.6
0.60
Appendix
Table 21: DiMoBench end-of-dialogue statistics per persona (350 turns; mean over three seeds). The means reproduce the DiMoBench row of Table 19 .
LoCoMo (gpt-4o-mini)
DiMoBench (350 t. end)
Config
Single
Multi
Temp.
Open
Overall
Nrawend
Acc. (%)
Nrawend
Neffend
A0 (default; no FACT prune)
81.41
68.49
66.21
39.63
73.57
287
84.0
62
27.6
A1 (FACT prune w<0.05 )
81.36
68.42
66.18
39.55
73.51
168
83.8
41
27.5
A2 (FACT prune w<0.10 )
80.95
68.05
65.74
39.21
73.06
142
82.9
33
26.7
Appendix
Table 22: FACT pruning ablation. A1 ( w<0.05 ) shrinks raw memory by 41% on LoCoMo and 34% on DiMoBench while losing ≤0.2 accuracy points, inside seed variance. A2 ( w<0.10 ) starts pruning items that still have residual influence and degrades accuracy by 0.5 – 1.1 points.
Benchmark
Recall@10
QA (hit)
QA (miss)
QA (forced hit)
Gap recovered
DiMoBench
95.0 ± 0.4
84.8
64.2
70.7
32%
DiMoBench - XL
93.0 ± 0.2
84.7
63.0
69.4
29%
LongMemEval - S KU
90.0 ± 0.5
93.4
75.8
80.3
26%
AMemGym
89.1 ± 0.5
72.1
55.8
60.6
29%
Appendix
Table 23: Belief-update recall and the cost of a miss. Recall@10 of the belief due for update (mean ± std over three seeds); question accuracy (%) on events whose target was retrieved (hit) or not (miss), and on missed events after forcing the target into the Top-10 (forced hit); the last column is the share of the hit–miss gap recovered by the forced hit. AMemGym accuracies are changed-state probe accuracies.
Q1
Q2
Q3
Q4
Q5
Gold-evidence Precision@ K↑
0.58
0.68
0.76
0.82
0.87
Gold-evidence Recall@ K
0.86
0.88
0.89
0.86
0.82
Contradiction accuracy ↑
0.71
0.79
0.84
0.89
0.92
Update correctness ↑
0.74
0.80
0.85
0.90
0.93
False- CORRECT (%) ↓
2.1
1.0
0.8
0.7
0.6
Downstream probe accuracy ↑
0.66
0.71
0.75
0.78
0.81
Appendix
Table 24: Calibration of ρt on AMemGym (2,829 gate events over three seeds, split into ρt quintiles from Q1, lowest, to Q5, highest; means over seeds). Precision@ K and Recall@ K use different references (the user’s current gold state versus the single belief due for update); false- CORRECT is the share of CORRECT firings in the quintile judged unwarranted in the blind audit.
Track
Gold shifts
ρ -assessable
d=0
d=1
d=2
d>2
Any delay
DiMoBench
45
41
38
3
0
0
3/41=7.3%
DiMoBench - XL
135
122
104
12
3
3
18/122=14.8%
Appendix
Table 25: CORRECT delay at gold change points on DiMoBench and DiMoBench - XL (counts pooled over three seeds). A shift is ρ -assessable when the outdated belief is in the Top- K and Sraw>θhigh ; d is the delay in turns.
Method
Retrieval
End-to-End
P50
P95
P50
P95
MemoryOS
1.66s
2.21s
–
–
MemOS
1.69s
3.44s
2.04s
3.89s
A-Mem
1.74s
7.23s
2.09s
7.68s
Mem0
2.44s
4.29s
2.74s
4.69s
PreCoMem (Ours)
1.95s
3.86s
2.35s
4.36s
Appendix
Table 26: Retrieval and end-to-end latency comparison. Evaluated on LoCoMo with gpt-4o-mini . “Retrieval” is the memory read/scoring cost only; “E2E” is the full query-to-answer time (retrieval + generation). PreCoMem obtains competitive typical latency and a well-controlled tail at both stages, without the extreme P95 spikes seen for recursive agentic baselines.
Figure 10: Compact view of a Belief Store snapshot. Each entry shows the type, natural-language content, confidence weight, and pointers to the episodic cache. FACT items are committed, HYPOTHESIS items are tentative; confidences modulate retrieval priority but are never shown to the dialogue agent directly.
Figure 11: Qualitative analysis of evidence consolidation. Left: In skill calibration, PreCoMem weighs credentials against contradictory behavioral cues to avoid over-estimation. Right: In project reasoning, it distinguishes high-confidence competencies from tentative learning (“loading”), preventing premature recommendations.
Large language model (LLM) agents increasingly rely on external memory to support long-horizon interaction, personalized assistance, and multi-step reasoning. However, existing memory systems still face three core challenges: they often rely too heavily on semantic similarity, which can miss evidence crucial for user-centric understanding; they frequently store related experiences as isolated fragments, weakening temporal and causal coherence; and they typically use static memory granularities that do not adapt well to the requirements of different questions. We propose AdaMem, an adaptive user-centric memory framework for long-horizon dialogue agents. AdaMem organizes dialogue history into working, episodic, persona, and graph memories, enabling the system to preserve recent context, structured long-term experiences, stable user traits, and relation-aware connections within a unified framework. At inference time, AdaMem first resolves the target participant, then builds a question-conditioned retrieval route that combines semantic retrieval with relation-aware graph expansion only when needed, and finally produces the answer through a role-specialized pipeline for evidence synthesis and response generation. We evaluate AdaMem on the LoCoMo and PERSONAMEM benchmarks for long-horizon reasoning and user modeling. Experimental results show that AdaMem achieves state-of-the-art performance on both benchmarks. The code will be released upon acceptance.
We investigate when belief-based memory actually improves large language model (LLM) agents. Our vehicle is Nous, a long-term memory architecture that represents each entity-attribute pair as a categorical probability distribution updated through closed-form Bayesian inference, with information-theoretic surprise driving belief revision and entropy-based forgetting. A controlled ablation on the LoCoMo benchmark shows that Bayesian belief updating alone provides little benefit over naive last-write-wins because existing conversational memory benchmarks rarely contain contradictory or differently reliable evidence. We then introduce reliability-conditioned updating, estimating per-observation reliability from epistemic language, and show on a controlled contradiction benchmark that belief updating substantially outperforms last-write-wins and raw-memory retrieval when observations differ in trustworthiness. Because content-derived reliability is itself vulnerable to manipulation, we further propose provenance-capped belief updating, where trust is bounded by source provenance rather than textual confidence. Under controlled memory-poisoning experiments, this approach resists volumetric poisoning attacks while revealing the utility costs and implementation requirements of provenance-aware memory. Finally, we quantify a 27.5-point discrepancy between strict token-F1 and LLM-as-judge evaluation on identical outputs, highlighting important reproducibility concerns for long-term memory benchmarks. Our results suggest that probabilistic belief-based memory is most beneficial in environments requiring reasoning over conflicting and differently trustworthy evidence, rather than conventional conversational recall alone.
Large Language Models (LLMs) serving as long-term dialogue agents require memory systems that support reliable reasoning over extended interactions. However, existing Retrieval-Augmented Generation (RAG) frameworks typically treat memory as passive storage, making it difficult to distinguish source-attributed beliefs from unattributed event/fact records and to connect evidence dispersed across sessions. We introduce CogMem, a cognitive memory architecture based on the PEC2F (Person-Event-Concept-Claim-Fact) graph schema. Dedicated Claim nodes preserve the source and target of subjective statements, while Fact and Event nodes represent semantic and episodic knowledge. Dialogue turns are incrementally converted into provenance-aware graph records, consolidated into higher-level facts, and reconciled into temporally scoped Claim views when the same source provides conflicting updates. For retrieval, a rule-based controller driven by LLM intent parsing composes four deterministic graph operators---anchoring, traversal, intersection, and evidence grounding---to reconstruct query-relevant context. Experiments on LoCoMo and LongMemEval show strong performance, especially on multi-hop, temporal, and knowledge-update tasks. Ablations and a semantic-collapse probe support complementary contributions from epistemic separation, consolidation, and agentic retrieval. Code: https://github.com/Silent-Rain02/CogMem.
Zirui Liao, Zhengxian Wu, Zhuohong Chen +4
Tsinghua University Shenzhen International Graduate School, Shenzhen, China