Long-horizon dialogue agents require effective memory systems to recall facts, events, and relationships from interaction histories. In this work, we propose a hybrid-memory architecture comprising episodic memories and structured relational memories and a consolidation mechanism which preserves efficiency. We conduct an empirical study across two benchmarks: LoCoMo and LongMemEval-S. Our experiments demonstrate that the two memory types are consistently complementary, combining both outperform state-of-the-art memory frameworks and achieves 90.2% accuracy on LoCoMo and 85.5% on LongMemEval-S.
Figures & tables
Figure 1: Architecture Overview. Stage 1 (Memory Curation): The Episodic Memory Manager stores timestamped summaries, while the Relational Memory Manager extracts structured knowledge triplets with conversational context. Stage 2 (Consolidation): Memory entries are clustered based on semantic similarity, with memories merged or deleted to maintain efficiency. Stage 3 (Memory Retrieval and QA): Given a user query, relevant memories are retrieved from both episodic and relational stores to generate contextually grounded responses.
Model
Method
Single-Hop
Multi-Hop
Open Domain
Temporal
Overall
F1
BLEU
LLM Score
F1
BLEU
LLM Score
F1
BLEU
LLM Score
F1
BLEU
LLM Score
F1
BLEU
LLM Score
gpt-4o-mini
FullContext
0.294
0.220
0.804
0.247
0.168
0.532
0.192
0.146
0.604
0.386
0.269
0.903
0.328
0.232
0.790
LangMem
0.388
0.331
0.614
0.335
0.239
0.524
0.294
0.235
0.476
0.319
0.262
0.249
0.358
0.294
0.513
Mem0
0.444
0.377
0.681
0.343
0.252
0.603
0.271
0.194
0.406
0.444
0.376
0.504
0.415
0.342
0.613
RAG
0.222
0.186
0.320
0.186
0.117
0.313
0.190
0.135
0.326
0.195
0.157
0.237
0.208
0.164
0.302
Zep
0.397
0.337
0.632
0.275
0.193
0.505
0.229
0.157
0.396
0.448
0.381
0.589
0.375
0.309
0.585
Table 1: Performance on the LoCoMo dataset using LLM-as-a-judge, F1 and BLEU scores. Baselines LangMem, Mem0, RAG, Zep and Nemori were taken from Nan et al. (2025) . Bold indicates the best among memory methods.
Question Type
Full-context
Nemori
Ours
(101K tokens)
(3.7–4.8K tokens)
(3–6.9K tokens)
single-session-preference
0.167
0.867
0.500
single-session-assistant
0.982
0.929
0.982
temporal-reasoning
0.602
0.722
0.858
multi-session
0.511
0.556
0.801
knowledge-update
0.769
0.795
0.902
Table 2: Detailed performance comparison on LongMemEval-S (gpt-4.1-mini) across question types.
Method
LLM Score
Tokens
FullContext
0.790
21,788
RAG-4096
0.302
3,430
LangMem
0.513
125
Zep
0.585
2,247
Mem0
0.613
1,027
Nemori
0.744
2,745
Table 3: Performance and efficiency comparison on LoCoMo dataset with gpt-4o-mini.
Model
Overall
S-Hop
M-Hop
Open
Temp.
gpt-5.4
0.906
0.865
0.903
0.729
0.941
gpt-4.1-mini
0.902
0.858
0.872
0.697
0.952
gpt-4o-mini
0.835
0.765
0.813
0.645
0.888
Qwen3-8B
0.762
0.748
0.629
0.552
0.841
Table 4: Performance of our framework with different backbone LLMs on LoCoMo.
Figure 2: Effect of retrieval budget k on LLM Score with gpt-4.1-mini . Performance plateaus at k=10 .
mcs
Compression
LLM Score
Clusters
2
30.9%
0.876
2208
5
21.7%
0.896
693
10
16.3%
0.892
325
20
9.6%
0.898
144
Table 5: Effect of min_cluster_size on consolidation.
Appendix figures & tables3 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 8Figure 9
Component
Overall
S-Hop
M-Hop
Open
Temp.
gpt-4o-mini
Episodic only
0.776
0.765
0.710
0.645
0.820
Relational only
0.730
0.680
0.750
0.572
0.757
Ours
0.835
0.765
0.813
0.645
0.888
+ Consolidation
0.823
0.797
0.764
0.625
0.875
gpt-4.1-mini
Appendix
Table 6: Memory component ablation on LoCoMo with gpt-4o-mini and gpt-4.1-mini .