Long-term memory enables large language model (LLM) agents to leverage historical interactions for future tasks. However, existing memory systems struggle to utilize continuously evolving historical information, as they often rely on static memory representations and single-round retrieval strategies, failing to track factual changes or integrate distributed evidence across long-term interactions. To address these challenges, we propose EMIR2, an Evolution-Aware Memory framework with Intent-Guided Multi-Round Retrieval, enabling LLM agents to maintain evolving historical knowledge and adaptively retrieve relevant evidence. Specifically, EMIR2 constructs a State-Evolving Memory Graph (SEMG) that represents long-term memory as evolving knowledge states supported by temporal event trajectories and evidential associations. By maintaining semantic states through evidence-based updates, SEMG preserves historical evolution and enables evidence tracing under complex and conflicting scenarios. Building upon this, we introduce an intent-guided multi-round retrieval mechanism that iteratively identifies missing evidence and expands retrieval based on accumulated information. Experiments on LoCoMo and MemConflict demonstrate that EMIR2 improves long-term memory utilization, dynamic and static conflict handling, and complex retrieval performance, achieving relative improvements of more than 12% in certain categories. These results highlight the effectiveness of jointly modeling memory evolution and adaptive evidence acquisition for long-term agent interactions.
Figures & tables
Figure 1: Illustrative examples of three types of memory conflicts in long-term interactions.
Figure 2: Overview of EMIR 2 . (a) Episodic units are organized into temporal chains while maintaining evolving semantic states to construct the SEMG. (b) The LLM controller adaptively selects retrieval actions based on the query intent and accumulated evidence until sufficient information is acquired.
Model
Method
Multi-hop
Temporal
Open Domain
Single hop
Overall
Gemini 2.5 Flash
RAG
58.16 ±0.45
49.22 ±0.15
41.67 ±0.47
69.20 ±0.05
61.30
A-Mem ( Xu et al., 2025 )
53.54 ±0.33
49.53 ±0.44
33.33 ±0.49
61.83 ±0.10
55.97
MemoryOS ( Kang et al., 2025 )
63.82 ±0.44
47.04 ±0.53
41.66 ±0.85
71.90 ±0.22
63.35
LangMem ( LangChain, 2025 )
61.34 ±0.79
53.58 ±0.25
38.54 ±0.33
69.68 ±0.12
62.86
Mem0 ( Chhikara et al., 2025 )
68.79 ±1.21
61.68 ±0.29
41.66 ±1.63
73.72 ±0.10
68.31
MRAgent ( Ji et al., 2026 )
75.17 ±0.33
80.37 ±0.15
68.75 ±0.98
90.48 ±0.10
84.21
Table 1: LLM-Judge performance comparison of different memory methods across various LLM backbones. All values are reported as percentages. Results are averaged over three independent runs, with standard deviations reported as ± values. Bold numbers indicate the best performance among all methods, while underlined numbers indicate the best performance among baseline methods.
Figure 3: Ablation study of EMIR 2 on the LoCoMo benchmark. All experiments use Gemini 2.5 Flash for memory construction and answer generation, Qwen3.5-9B for retrieval controller, and GPT-4o-mini as the LLM judge. We select the best result from the main experiments as the score for the Full model.
Method
Dynamic
Static
Conditional
Average
AA ↑
SEH@3 ↑
AA ↑
SEH@3 ↑
AA ↑
SEH@3 ↑
AA ↑
A-Mem
0.3596
0.5205
0.2639
0.3611
0.7122
0.8111
0.4452
LangMem
0.4966
0.7842
0.1944
0.3194
0.1556
0.2012
0.2822
Letta
0.3955
0.5394
0.2223
0.4167
0.8435
0.9046
0.4871
MemOS
0.3793
0.5548
0.4375
0.5694
0.8449
0.8889
0.5539
Mem0
0.1224
0.2003
0.1944
0.2917
0.7667
0.8222
0.3612
Table 2: Conflict-aware memory evaluation on MemConflict. Bold numbers indicate the best performance among all methods, while underlined numbers indicate the best performance among baseline methods.
Figure 4: Analysis of adaptive multi-round retrieval. (a) Performance changes with increasing retrieval rounds. (b) Accuracy-cost trade-off under different retrieval budgets.
Appendix figures & tables8 assets
Supplementary material from the paper’s appendix.
Appendix
Category
Relation
Semantics
Structural
FIRST_ TIMELINE_NODE
Chain head → Earliest event Provides entry from a topic-chain head to its earliest episodic event.
Structural
TEMPORAL_NEXT
Earlier event → Later event Connects consecutive events in chronological order. It expresses temporal adjacency, not causality.
Structural
INITIAL_STATE
Chain head → Initial semantic state Provides entry to the earliest separately represented semantic state, when such a node exists.
Structural
CURRENT_STATE
Chain head → Current semantic state Provides direct access to the currently accepted semantic state of a topic chain.
Structural
SEMANTIC_NEXT
Earlier state → Successor state Connects consecutive semantic epochs in their order of evolution.
Structural
SUPPORTED_BY
Event or state → Chain head Attaches an event or semantic state to its topic chain for structural organization and navigation.
Appendix
Table 3: Complete SEMG relation taxonomy. Structural relations encode graph organization, while event and state relations capture semantic connections supported by extracted evidence. Both relation types support graph traversal and retrieval.
Category
Qwen3.5-9B
Claude Controller
MRAgent
Multi-hop
88.53
91.49
90.19
Temporal
85.25
87.54
85.34
Appendix
Table 4: Effect of retrieval controller capacity under the Claude Sonnet 4.5 backbone. All configurations of EMIR 2 are kept unchanged except for the retrieval controller. Bold numbers indicate the best performance within each category.
Category
I → C
C → I
C → C
I → I
Multi-hop
16
8
242
16
Temporal
15
8
266
32
Overall
31
16
508
48
Appendix
Table 5: Instance-level prediction transitions between the Qwen3.5-9B and Claude Sonnet 4.5 retrieval controllers, where I and C denote incorrect and correct predictions, respectively.
Category
Avg. Rounds
Decomposition Rate
Complete Gold Rate
Complete Gold
Multi-hop
Qwen3.5-9B
3.58
14.5%
43.6%
Multi-hop
Claude Sonnet 4.5
2.55
42.6%
47.5%
Temporal
Qwen3.5-9B
3.22
3.7%
88.2%
Temporal
Claude Sonnet 4.5
2.51
12.8%
91.0%
Appendix
Table 6: Retrieval behavior comparison between the Qwen3.5-9B and Claude Sonnet 4.5 controllers. Complete Gold denotes the proportion of questions for which all annotated supporting evidence is retained.
Retrieval Statistics
Full
w/o cross-rel
Graph Expansion Count
5957
5548
Candidates from Graph Expansion
3503
2800
Final Candidates from Graph Expansion
2372
1798
Final Retrieved Candidates
18083
17997
Appendix
Table 7: Retrieval behavior analysis of the full model and the w/o cross-rel variant.
Evidence Statistics
Full
w/o cross-rel
Retrieved Gold Evidence
1721
1723
Retrieved via Graph Expansion
197
101
Uniquely Retrieved Evidence
41
36
Appendix
Table 8: Analysis of evidence retrieved through cross-chain edges.
Method
Dynamic
Static
Conditional
Average
AA
UOCS
AA
CRS
AA
AA
A-Mem
0.3596
0.2911
0.2639
0.2501
0.7122
0.4452
LangMem
0.4966
0.3579
0.1944
0.2083
0.1556
0.2822
Letta
0.3955
0.3527
0.2223
0.2031
0.8435
0.4871
MemOS
0.3793
0.3818
0.4375
0.2361
0.8449
0.5539
Mem0
0.1224
0.1130
0.1944
0.1528
0.7667
0.3612
Appendix
Table 9: Black-box performance of memory systems on MemConflict by conflict type. Bold numbers indicate the best performance among all methods, while underlined numbers indicate the best performance among baseline methods.
Method
Dynamic
Static
Conditional
Average
SEH@3
SRS
SEH@3
SRS
SEH@3
SRS
SEH@3
SRS
A-Mem
0.5205
0.4341
0.3611
0.2854
0.8111
0.7288
0.5642
0.4828
LangMem
0.7842
0.7089
0.3194
0.2697
0.2012
0.1944
0.4349
0.3910
Letta
0.5394
0.4620
0.4167
0.3099
0.9046
0.7653
0.6202
0.5124
MemOS
0.5548
0.4552
0.5694
0.4886
0.8889
0.8198
0.6710
0.5879
Mem0
0.2003
0.1587
0.2917
0.2401
0.8222
0.7780
0.4381
0.3923
Appendix
Table 10: White-box memory retrieval and ranking of memory systems on MemConflict by conflict type. Bold numbers indicate the best performance among all methods, while underlined numbers indicate the best performance among baseline methods.
Long-term memory is essential for LLM-based agents to sustain interactions and reliably leverage distant history. However, existing memory systems typically process heterogeneous dialogue content through a uniform summarization and retrieval pipeline, leading to either excessive token consumption or irreversible loss of fine-grained evidence. We argue that historical dialogue content should be handled differently according to its compressibility, temporal dynamics, and fidelity requirements. Based on this insight, we propose LeanMem, a lightweight long-term memory framework. LeanMem first filters out low-value content, then stores informative segments as compact profile memory, temporally structured event memory, or source-grounded record memory, depending on the nature of the information. During maintenance, only dynamically evolving event memories are selectively updated, avoiding redundant consolidation of stable profiles and immutable records. During inference, LeanMem dynamically selects memory types and allocates retrieval budgets according to query-specific evidence demands, assembling relevant evidence on demand. On LoCoMo and LongMemEval-S with GPT-4.1-mini and Qwen3-8B, LeanMem improves accuracy over the strongest memory-based baseline in every setting, by up to 15.1 points, at the lowest or near-lowest construction cost, inference tokens, and latency. The code and datasets are included in the supplementary materials.
Long-term LLM agents need persistent memory that can track changing facts and provide relevant evidence across sessions. Existing memory systems often store observations as isolated records, summaries, or indexed fragments, which makes evidence aggregation, fact revision, and memory maintenance difficult. We propose Infini Memory, a maintainable text-based persistent memory architecture that treats agent memory as topic-structured documents. Each topic document serves as a semantic unit for collecting related evidence, preserving metadata, and revising facts over time. New observations are first staged in a buffer and periodically consolidated into coherent textual contexts. At inference time, an agentic retrieval procedure lets the LLM read memory through iterative tool calls rather than a single retrieval step. On MemoryAgentBench, Infini Memory achieves 64.7% overall score. Ablations show that topic-structured maintenance and iterative evidence inspection improve complementary aspects of long-term memory use.
Suozhao Ji, Baodong Wu, Zehao Wang +8
1Infinigence AI · 2Tsinghua University · 3Shanghai Jiaotong University
Long-term memory is essential for LLM-based agents operating over extended interactions. Existing memory systems primarily update memory when new information arrives, treating retrieval as the endpoint of memory access rather than a driver of memory evolution. Consequently, retrieval feedback is rarely exploited to reorganize memory for future access continuously. Moreover, most existing approaches rely on predefined memory structures together with fixed retrieval pipelines, limiting the agent's ability to organize and evolve its own memory autonomously. Inspired by memory reconsolidation in cognitive neuroscience, we propose REALM, a reconsolidation-evolution agentic long-term memory framework. It models long-term memory as a continual lifecycle by autonomously organizing memories into a heterogeneous cognitive graph, retrieving evidence via adaptively composed graph-search atoms, and continually reconsolidating memories based on retrieval feedback. REALM achieves an average accuracy of 75.97% on LoCoMo and 65.11% on LongMemEval, outperforming the strongest baselines by 7.17 and 1.31 points respectively. Ablation studies confirm that memory reconsolidation consistently boosts performance, with further analyses revealing that it progressively reorganizes related memory units into more coherent local structures for collective evidence recall and utilization during reasoning. These results suggest that retrieval-driven memory reconsolidation provides an effective mechanism for continually evolving long-term memory in LLM agents.
Yuanyi Song, Yukai Wang, Xinbei Ma +7
Shanghai Jiao Tong University · OPPO · National University of Singapore