FTA-Mem: Fact-Time-Affect Anchored Memory for Low-Density Long-Term Dialogue

Chang Liu; Junyi Zhao; Shuyi Zhang; Changsheng Ma; Yongfeng Tao; Minqiang Yang; Bin Hu

cs.CLAug 17, 2026

FTA-Mem: Fact-Time-Affect Anchored Memory for Low-Density Long-Term Dialogue

Authors: Chang Liu, Junyi Zhao, Shuyi Zhang, Changsheng Ma, Yongfeng Tao, Minqiang Yang, Bin Hu

Organizations: School of Information Science and Engineering, Lanzhou University

Abstract

Long-term emotional-support agents require memory mechanisms for personalized understanding across sessions. However, emotional-support dialogue is often low-density: turns are incomplete, evidence is scattered, and user states evolve over time. Existing memory methods usually rely on fixed units, such as turn-level notes or session summaries, which may lose details or introduce redundant noise. We propose FTA-Mem, a structured memory framework for low-density long-term dialogue. FTA-Mem uses Boundary-preserving Window Segmentation (BWS) to form coherent situation fragments, and constructs Fact-Time-Affect Memory Units (FTA Units) that jointly encode factual content, temporal grounding, and affective context. Retrieved units are then synthesized into structured context for answer generation. Experiments on ES-MemEval and LoCoMo show that FTA-Mem improves overall long-term memory question answering across benchmarks with different information-density characteristics. On ES-MemEval, FTA-Mem achieves 0.3871 F1 and 0.6668 BERTScore. Further analysis shows that situation-level FTA construction better balances evidence preservation and construction cost than coarse session-level or overly fine-grained turn-pair construction, providing an effective granularity trade-off for long-term dialogue memory.

Figures & tables

Figure 1: Overview of FTA-Mem. BWS segments dialogue turns into boundary-preserving fragments, which are converted into Fact-Time-Affect units. Partial units can be carried forward, adjacent units can be fused before ID assignment, and finalized units are linked, retrieved, deduplicated, and organized into structured context for grounded answer generation.

Metric	ES-MemEval	LoCoMo
Turn length	18.56	22.69
Factual anchors / turn	0.80	1.11
Entity anchors / turn	0.52	1.24
Temporal anchors / turn	0.11	0.22
Support-process anchors / turn	0.91	0.50
Implicitness (1–3)	1.91	1.58

Table 1: Turn-level information-density statistics of ES-MemEval and LoCoMo. Values are averaged over turns.

Backbone	Method	F1 Score (%) $\uparrow$						BERTScore (%) $\uparrow$						Judge $\uparrow$	Avg.R $\downarrow$
Backbone	Method	IE	TR	CD	Abs	UM	All	IE	TR	CD	Abs	UM	All	Judge $\uparrow$	Avg.R $\downarrow$
Qwen3-8B	MemoryBank	23.13	29.76	25.65	33.29	24.55	27.08	53.50	64.45	59.95	49.37	65.31	58.66	1.036	4.3
	MemGPT	44.15	32.77	25.14	29.54	25.64	31.69	65.56	65.28	58.01	49.26	65.93	61.19	1.107	2.8
	MemoryOS	21.41	25.79	22.35	59.96	22.29	29.70	52.35	60.83	56.73	67.48	63.87	60.09	0.917	4.9
	A-Mem	27.96	31.66	26.60	39.52	25.22	29.97	56.18	65.22	60.12	53.46	65.55	60.23	1.090	3.0
	CompassMem	22.06	28.85	22.06	56.52	24.32	30.20	54.09	62.35	55.69	66.98	64.70	60.67	1.024	4.4

Table 2: Performance on ES-MemEval. Avg.R is computed over the twelve automatic metrics, i.e., F1 and BERTScore on IE/TR/CD/Abs/UM/All; Judge is not included in Avg.R. Best results are in bold, and second-best results are underlined.

Backbone	Method	Single Hop		Multi Hop		Temporal		Open Domain		Average $\uparrow$
Backbone	Method	F1 $\uparrow$	BLEU-1 $\uparrow$	F1 $\uparrow$	BLEU-1 $\uparrow$	F1 $\uparrow$	BLEU-1 $\uparrow$	F1 $\uparrow$	BLEU-1 $\uparrow$	F1	BLEU-1
Qwen3-8B	MemoryBank	14.32	16.07	28.30	23.78	22.25	18.69	15.32	12.33	20.05	17.72
	MemGPT	21.66	22.43	19.14	15.28	3.71	2.81	11.86	9.28	14.09	12.45
	MemoryOS	18.65	22.67	26.07	22.31	3.62	2.52	14.21	11.44	15.64	14.74
	A-Mem	23.92	22.69	29.42	24.85	26.81	21.85	15.68	11.90	23.96	20.32
	CompassMem	31.73	29.12	42.05	35.86	34.18	28.06	18.64	14.67	31.65	26.93

Table 3: Results on LoCoMo category-level long-term memory question answering. Best results are in bold, and second-best results are underlined.

Gain	Fact	Ent.	Time	Impl.
ES-MemEval
FTA $-$ CompassMem	-0.39	-0.12	-0.14	0.41
FTA $-$ Avg.	-0.32	-0.07	-0.08	0.44
LoCoMo
FTA $-$ CompassMem	0.00	0.47	-0.44	0.21
FTA $-$ Avg.	0.26	0.40	-0.12	-0.10

Table 4: Diagnostic Pearson correlations between dialogue-level density fields and FTA-Mem’s relative F1 gain. Fact, entity, and time densities are mean annotated anchors per turn. Avg. denotes the mean over MemoryBank, MemGPT, MemoryOS, A-Mem, and CompassMem.

Method	Retrieval Unit	Units	Token/Q
ES-MemEval
MemoryBank	Summary/profile	838	982
MemGPT	Archival passage	401	3,003
MemoryOS	Turn-level state	4,321	3,658
A-Mem	Agentic note	401	3,103
CompassMem	Event graph node	2,642	1,268

Table 5: Answer-stage cost comparison. Token/Q denotes response-generation tokens per question.

Unit	F1 $\uparrow$	B/B1 $\uparrow$	Units	Calls	Tokens
ES-MemEval
Session-level	31.76	61.78	1,692	401	1.58M
Turn-pair	37.06	65.43	8,556	4,883	6.40M
FTA-Mem	38.71	66.68	6,564	3,245	4.99M
LoCoMo
Session-level	29.34	25.13	1,008	272	1.09M

Table 6: Granularity and memory-unit construction cost. Units is the total number of constructed memory units. Calls and Tokens count only LLM calls and tokens used for memory-unit construction.

Figure 2: Effect of retrieval budget K on ES-MemEval and LoCoMo.

Figure 3: Ablation results on ES-MemEval and LoCoMo.

Dataset	Setting	F1 $\uparrow$	B/B1 $\uparrow$	Units
ES-MemEval	BWS	38.71	66.68	6,564
ES-MemEval	Fixed Window	37.94	65.34	6,678
LoCoMo	BWS	37.35	31.67	4,299
LoCoMo	Fixed Window	37.04	30.84	4,332

Table 7: Comparison between boundary-preserving window segmentation and fixed-window segmentation. B/B1 denotes BERTScore on ES-MemEval and BLEU-1 on LoCoMo. Scores are reported in percentages.

Setting	F1 $\uparrow$	BERTScore $\uparrow$
FTA-Mem	38.71	66.68
w/o Query Rewrite	37.44	65.90
Plain Embedding	35.99	64.88
Flat Context	37.86	65.24
w/o Auxiliary Memory	38.65	66.73

Table 8: Additional ablation results on ES-MemEval. Scores are reported in percentages.

Dataset	Gain	Field	Pearson $r$	Spearman $\rho$	95% CI	Perm. $p$	LOO Range
ES-MemEval	FTA–CompassMem	Fact	-0.39	-0.48	[-0.70, -0.04]	0.110	[-0.49, -0.32]
ES-MemEval	FTA–CompassMem	Implicit.	0.43	0.41	[0.00, 0.73]	0.077	[0.31, 0.53]
ES-MemEval	FTA–Avg. Baseline	Fact	-0.32	-0.36	[-0.74, 0.13]	0.154	[-0.54, -0.20]
ES-MemEval	FTA–Avg. Baseline	Implicit.	0.44	0.34	[-0.15, 0.78]	0.081	[0.23, 0.55]
LoCoMo	FTA–CompassMem	Fact	0.00	-0.20	[-0.77, 0.60]	0.481	[-0.53, 0.07]
LoCoMo	FTA–CompassMem	Implicit.	0.21	0.31	[-0.16, 0.96]	0.306	[0.26, 0.63]

Table 9: Additional density-gain correlations under the same setting as the main-paper diagnostic table. Pearson

r

follows the main-paper aggregation. CI denotes case-level bootstrap 95% confidence interval; Perm.

p

is computed by permutation testing; and LOO Range reports leave-one-case-out Pearson ranges.

Figure 4: Dialogue-level density-gain trends on ES-MemEval.

Dataset	Clusters	Mean Diff.	95% CI
ES-MemEval	18	8.43	[5.81, 11.04]
LoCoMo	10	0.55	[-0.01, 3.18]

Table 10: Paired cluster bootstrap uncertainty check for FTA-Mem minus CompassMem. Values are token-level F1 differences in percentage points. Clusters correspond to ES-MemEval cases or LoCoMo conversations. LoCoMo excludes Cat5 no-information questions.

Benchmark	Fact	Time	Affect/Ctx.	Rel.
ES-MemEval	2.00	2.00	1.48	1.36
LoCoMo	2.00	2.00	1.38	1.25

Table 11: Manual audit of intermediate FTA memory quality. Scores are reported on a 0–2 scale. Rel. denotes relation relatedness for linked memory pairs.

Query	Gold Answer	A-Mem	Ours
Q1: As of October 12, 2025 , what event made Sarah feel unsettled? ( Temporal Reasoning )	An argument with Jack about future plans .	Sarah felt unsettled due to her boss’s unexpected resignation and concerns about leadership changes.	Sarah felt unsettled after an argument with Jack about their future plans .
Q2: Has Sarah found time for painting consistently as a coping strategy? ( Conflict Detection )	No , she has found it hard to find time for painting due to work and personal issues .	Yes , Sarah has been painting consistently as a coping strategy.	Sarah finds painting calming, but struggles to find time due to work and personal issues .
Q3: What future plan indicates Jimmy’s willingness to reconnect with past relationships ? ( User Modeling )	He is considering reaching out to George for an honest conversation.	Jimmy’s willingness to reconnect is indicated by his desire to meet with Jennifer .	Jimmy plans to reach out to George for an honest conversation.

Table 12: Case study on ES-MemEval. A-Mem often retrieves related but noisy memories, while FTA-Mem better identifies the target situation, temporal state, and participant relation.

Explore similar work

LeanMem: Simple and Efficient Long-Term Memory for LLM Agents
Aug 4, 2026Yuxin Liao, Le Wu, Min Hou +3Long-Term Memory Simple
Hefei University of Technology
LazyMem: Retrieve Broadly, Construct Selectively for Efficient Long-Term Agent Memory
Jul 17, 2026Jing Yu, Yibo Zhao, Jiaming Zhang +1Selective
School of Data Science and Engineering, East China Normal University
JustMem: Just-Enough Memory Access for Long-Term Conversations
Sep 17, 2026Guanhua Chen, Yanting Wang, Wenjing Zhi +1Conversational Memory Peak Memory
Beihang University

Aug 4, 2026, cs.AIJ/K move · Enter open · S save

LeanMem: Simple and Efficient Long-Term Memory for LLM Agents

Yuxin Liao, Le Wu, Min Hou, Hao Liu+2

Hefei University of Technology

Long-Term Memory Simple

Open paper AskPDF arXiv

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.

Abstract

Explore similar work

LeanMem: Simple and Efficient Long-Term Memory for LLM Agents

LazyMem: Retrieve Broadly, Construct Selectively for Efficient Long-Term Agent Memory

JustMem: Just-Enough Memory Access for Long-Term Conversations