Recent advances in persona-centric memory have revealed the powerful capability of multi-agent systems in managing persona memory, especially in conversational scenarios. However, these complex frameworks often suffer from information loss and are fragile across varying scenarios, resulting in suboptimal performance. In this paper, we propose DeltaMem, an agentic memory management system that formulates persona-centric memory management as an end-to-end task within a single-agent setting. To further improve the performance of our agentic memory manager, we draw inspiration from the evolution of human memory and synthesize a user-assistant dialogue dataset along with corresponding operation-level memory updating labels. Building on this, we introduce a novel Memory-based Levenshtein Distance to formalize the memory updating reward, and propose a tailored reinforcement learning framework to further enhance the management capabilities of DeltaMem. Extensive experiments show that both training-free and RL-trained DeltaMem outperform all product-level baselines across diverse long-term memory benchmarks, including LoCoMo, HaluMem, and PersonaMem.
Figures & tables
Figure 1: i) Framework of DeltaMem. When ingesting a new session k , the memory manager will interact with the current memory state and finally generate a set of operations to update the current memory state. ii) Brief illustration of memory evolving guided data synthesis. iii) The reward modeling module of reinforcement learning via state transition.
Single Hop
Multi Hop
Open Domain
Temperal
Overall
Method
BS
F1
LJ
BS
F1
LJ
BS
F1
LJ
BS
F1
LJ
BS
F1
LJ
Training-free Multi-Agent Memory Manager
A-Mem
27.58
33.34
54.05
14.90
20.76
39.79
8.81
9.22
18.85
31.08
35.40
49.91
24.82
29.96
48.38
Mem0
35.34
43.17
60.88
22.27
33.99
52.84
18.19
23.34
40.62
44.53
54.21
59.50
33.79
42.55
57.86
LangMem
33.63
40.91
71.12
26.86
35.51
62.23
2 2.32
26.04
47.92
25.84
30.75
23.43
30.06
36.88
58.11
Zep
38.92
49.56
76.60
23.30
35.74
61.70
14.82
19.37
41.35
34.53
42.00
49.31
33.64
43.57
65.99
Table 1: Performance of different methods on LoCoMo. We use GPT-4o-mini as the answer model for all methods. As for the memory agent, all training-free baselines are evaluated with GPT-4o-mini. Note that Memory-R1 trains both the answer model and one module of its multi-agent system, and these two models are directly trained on LoCoMo. B1, F1, and LJ stand for BLEU-1 score, F1 score and LLM-as-a-Judge score, respectively.
Methods
Pref-Rec
Recall-Reason
New-Ideas
Pref-Evol
Recall-User
Gen-New
Recall-Facts
Overall
Training-free Multi-Agent Memory Manager
Mirix
30.91
63.64
17.20
48.92
34.11
19.30
41.18
38.37
Mem0
46.67
77.10
19.35
45.32
36.95
28.65
41.18
43.12
LightMem
37.58
82.15
14.70
50.36
48.58
14.62
43.14
44.82
Supermemory
54.55
79.12
26.88
54.44
53.65
52.30
54.90
53.88
Zep
60.61
80.47
3 6.92
52.52
57.88
51.46
56.86
56.71
Table 2: Performance of different methods on PersonaMem. Pref-Rec, Recall-Reason, New-Ideas, Pref-Evol, Recall-User, Gen-New, and Recall-Facts stand for providing preference-aligned recommendations, recalling the reasons behind previous updates, suggesting new ideas, tracking full preference evolution, recalling user-shared facts, generalizing to new scenarios, and recalling facts mentioned by the user, respectively. Following common settings, each question was evaluated by 3 separate runs.
Method
Memory Extraction
Memory Updating
Question Answering
Training-free Multi-Agent Memory Manager
Zep
-
47.28
55.47
Mem0
57.31
25.50
53.02
Mem0-Graph
57.85
24.50
54.66
Memobase
25.13
5.20
35.33
Supermemory
56.90
16.37
54.07
Table 3: Performance of different methods on HaluMem.
Figure 2: Reward dynamics during RL training of DeltaMem-4B-RL (left) and DeltaMem-8B-RL (right).
Figure 3: Ablation study on local lexical fidelity.
Threshold
0.00
0.50
0.60
0.70
0.75
0.80
0.90
B1
38.40
37.12
38.35
39.24
39.87
3 9.74
38.32
F1
48.92
47.93
48.75
50.37
5 0.72
50.78
48.75
Table 4: Impact of threshold τ on LoCoMo.
Figure 4: Average cumulative accuracy on HaluMem as the number of sessions increases.
Appendix figures & tables5 assets
Supplementary material from the paper’s appendix.
Appendix
Parameter
Value
Learning Rate
1e-6
Train Batch Size
64
Number of Training Epochs
1
Number of Rollouts
8
Rollout Temperature
1.0
Clip Ratio
0.2
Appendix
Table 5: Implementation details of RL training for DeltaMem.
Method
Memory Extraction
Memory Updating
Question Answering
Correct ↑
Hallucination ↓
Omission ↓
Correct ↑
Hallucination ↓
Omission ↓
Training-free Multi-Agent Memory Manager
Zep
-
47.28
0.42
52.31
55.47
21.92
22.62
Mem0
57.31
25.50
0.45
74.02
53.02
19.17
27.81
Mem0-Graph
57.85
24.50
0.26
75.24
54.66
19.28
26.06
Memobase
25.13
5.20
0.55
94.25
35.33
29.97
34.71
Appendix
Table 6: Full results of Halumem.
Methods
Mem0
DM-4B
DM-4B-RL
DM-8B
DM-8B-RL
QPM
1.00
0.00
0.13
0.02
0.13
Perf.
58.11
70.13
74.03
71.04
75.13
Appendix
Table 7: Number of queries per memory on LoCoMo. QPM stands for number of queries per memory, and DM is short for DeltaMem.
Figure 5: Average cumulative accuracy as a function of the number of sessions for LightMem and DeltaMem models with different model scales and training settings. The solid line shows the average accuracy over 20 users, and the shaded area indicates the variability across these users.