Long-term memory has been widely adopted in LLM-based agents to support persistent interaction and personalized assistance. Long-term multi-party interactions, however, accumulate substantially more detailed information, with decisions repeatedly updated and states evolving across participants and interaction stages. Temporal structure is therefore crucial for tracing such memories: relevant information often appears densely within short intervals, while transitions among different states are anchored by a strict temporal order. Existing temporally aware memory frameworks primarily store time as metadata or use temporal signals to guide retrieval over preconstructed memory structures, without using temporal structure to both recover locally concentrated evidence and reconstruct its complete event progression. To address these limitations, we propose T-Mem, a temporal-structure-based memory framework that preserves interactions and organizes query-relevant evidence through two temporal mechanisms. Temporal Evidence Discovery exploits the local concentration of related messages to recover overlooked evidence, while Temporal Memory Structuring models state transitions and supporting relations over time and compresses low-information spans within the resulting forest. Evidence Forest Linearization then serializes the structured evidence for answer generation. Experiments on two long-term multi-party memory benchmarks demonstrate that T-Mem consistently outperforms competitive memory frameworks across LLMs, parameter settings, domains, and query types.
Figures & tables
Figure 1: Overview of T-Mem . Metadata-expanded BM25 and Temporal Evidence Discovery recover candidate evidence; Temporal Memory Structuring links state evolution and support while compressing low-information spans; and Evidence Forest Linearization presents the resulting forest.
Methods
GroupMemBench
EverMemBench
Abstention
Multi- hop
Knowledge Update
Temporal
Term Ambiguity
User Implicit
Overall
Fine-Grained Recall
Memory Awareness
Profile Understanding
Overall
GPT-5.5
Embedding
54.68
54.95
23.36
55.56
39.62
55.10
48.32
43.04
87.78
41.96
63.25
BM25 (Expand)
62.58
57.69
20.56
53.08
38.67
53.06
49.26
52.89
84.87
52.31
67.38
A-Mem
56.12
46.15
31.78
35.80
33.01
61.22
42.82
41.73
88.14
50.46
64.92
HippoRAG2
72.66
56.59
35.51
51.85
52.83
48.97
54.50
57.48
89.52
52.68
71.04
Table 1: Main results on GroupMemBench and EverMemBench.
Figure 2: Key-evidence recall across retrieval depths.
Figure 3: Accuracy with flat versus forest-structured evidence.
Method
GPT-5.5
Deepseek-V4-Flash
T-Mem (Ours)
63.09
62.68
w/o Metadata Expansion
61.20 ( −1.89 )
60.40 ( −2.28 )
w/o Temporal Discovery
57.05 ( −6.04 )
56.37 ( −6.31 )
w/o Temporal Structuring
58.92 ( −4.17 )
57.45 ( −5.23 )
w/o Compression
60.54 ( −2.55 )
61.74 ( −0.94 )
Table 2: Ablation study across answer/judge backbones. Overall GroupMemBench accuracy (%) across 745 questions. Values in parentheses are absolute changes in percentage points relative to full T-Mem under the same backbone. For w/o Compression, contexts exceeding 8,192 tokens are directly truncated to the budget.
Figure 4: Robustness to temporally local revision noise on Knowledge Update. Complete Chain Recall measures recovery of all injected revisions and the final rollback; Revision w/o Rollback measures retrieval of intermediate revisions without the rollback (lower is better).
Figure 5: Overall GroupMemBench accuracy under restricted metadata availability. The dashed bar marks the strongest baseline.
Appendix figures & tables12 assets
Supplementary material from the paper’s appendix.
Appendix
Question type
Finance
Technology
Healthcare
Manufacturing
Total
Multi-hop
48
41
48
45
182
Knowledge update
32
36
17
22
107
Temporal
45
37
37
43
162
User implicit
15
28
2
4
49
Term ambiguity
45
43
7
11
106
Abstention
29
28
34
48
139
Appendix
Table 3: GroupMemBench question distribution used in our experiments.
Canonical field
GroupMemBench source
EverMemBench source
Adaptation and use
Message ID
msg_node
topic_id, date, group, message_index
EverMemBench has no native message-ID field; these four source fields form a stable provenance key that is never used as semantic evidence.
Content
content
dialogue
Original message text retained without summarization before retrieval.
Timestamp
timestamp
date, time
Normalized to an absolute time and used for ordering, clustering, and temporal constraints.
Channel/group
channel (top-level JSON key)
group (key in dialogues)
Defines the conversational stream within which messages are compared.
Phase
phase_name
Not provided
Included in the metadata-expanded message representation for GroupMemBench and omitted for EverMemBench.
Topic
topic
topic_id
Defines the topic-specific anchor partition.
Appendix
Table 4: Source metadata and field adaptation used for GroupMemBench and EverMemBench. Missing metadata is left empty rather than generated.
Method
Retrieval Top- k
Context cap (tokens)
Embedding
36
8,192
BM25
36
8,192
A-Mem
Variable
8,192
HippoRAG2
36
8,192
Hindsight
Variable
8,192
StructMem
60 atomic + 5 summaries
8,192
Appendix
Table 5: Baseline retrieval limits. Context caps apply to retrieved evidence, not to generated answers or judge outputs.
Method
Memory Processing Tokens
Answer and Judge Tokens
Total Tokens
Average Tokens per Question
Total Cost (USD)
Embedding
9.07M
5.80M
14.87M
19.96K
39.64
BM25 (Expand)
0
5.66M
5.66M
7.60K
38.68
A-Mem
392.08M
1.11M
393.19M
527.77K
76.70
HippoRAG2
129.57M
5.47M
135.04M
181.26K
65.50
Hindsight
481.67M
5.68M
487.35M
654.16K
122.59
StructMem
125.16M
8.33M
133.49M
179.18K
148.67
Appendix
Table 6: Efficiency on GroupMemBench. Token consumption is the primary efficiency measure, while total API-equivalent cost is reported as a complementary indicator for the complete 745-question evaluation. Embedding tokens are counted although the encoder can be deployed locally. The StructMem and MemGAS costs are estimated from stage-level token usage; downstream answer and judge calls use GPT-5.5-equivalent input/output pricing.
Figure 6: Analysis of Temporal Evidence Discovery and Temporal Memory Structuring. Left: key-evidence recall over 606 non-Abstention questions. T-Mem Reranked and Global Evidence Retrieval both output exactly 36 messages and share the same embedding, evidence-selection, and reranking models and configurations; the global control searches the entire topic instead of temporally localized intervals. The horizontal axis denotes the initial retrieval depth. Right: answer accuracy under flat, untyped-forest, and typed temporal-forest representations at different retrieval depths, using DeepSeek-V4.1-Flash as the answer and judge model.
Figure 7: Qualitative example of Temporal Evidence Discovery. BM25 and HippoRAG2 at Top-36 retrieve relevant messages but miss the state-changing message Msg_6416 . Temporal locality allows T-Mem to revisit the corresponding discussion episode, recover the overlooked update, and reconstruct the current four-channel review policy. Blue text highlights the answer-critical information recovered from the local temporal episode.
Figure 8: Qualitative example of Temporal Memory Structuring. An explicit Update chain distinguishes the earlier LMS-export policy, its challenge, and the revised append-only completion plan. Blue text marks the current update; content is condensed for readability.
Figure 9: Qualitative example of recovering a newly introduced constraint. Global retrieval reaches the correct field-tagging discussion but misses Msg_6579 , which changes the policy from a state tag alone to a two-part rule containing both display state and fallback eligibility. Temporal Evidence Discovery recovers this low-ranked update from the corresponding local episode. Blue text highlights the answer-critical constraint recovered through temporal locality.
Figure 10: Qualitative example combining Update and Expand relations. The temporal evidence forest distinguishes the revision of the control policy from the later message that adds implementation constraints. Update captures the transition to a clause-specific gate, whereas Expand attaches ownership and review requirements to the revised state. Blue text highlights the current policy and the relations used to organize its supporting evidence.
Figure 11: Answer and judge prompts for flat and temporal-forest evidence. Blue text denotes instructions present only in the forest interface; the judge prompt is shared.
Figure 12: Prompt for strict, query-conditioned evidence selection within a topic–time window.
Figure 13: Prompt for query-conditioned Temporal Memory Structuring.