AutoViewMem: Self-Configuring Orthogonal Views for Conversational Long-Term Memory

Zijie Cao; Xijun Qu; Zhicheng Gu; Xiaoshu Chen; Duanyang Yuan; Yanning Hou; Sihang Zhou; Jianxing Gong

cs.AISep 18, 2026

AutoViewMem: Self-Configuring Orthogonal Views for Conversational Long-Term Memory

Authors: Zijie Cao, Xijun Qu, Zhicheng Gu, Xiaoshu Chen, Duanyang Yuan, Yanning Hou, Sihang Zhou, Jianxing Gong,

Organizations: National University of Defense Technology, Changsha, China

Abstract

Long-term memory is essential for large language model (LLM) agents to maintain consistency and personalization over extended interactions. Existing memory systems typically rely on fixed granularities or static schemas, but these designs struggle when heterogeneous information, such as preferences, events, constraints, and temporal updates, is embedded in a single mixed representation. The resulting semantic interference makes top-K retrieval sensitive to noise and often leaves relevant evidence poorly ranked. We present AutoViewMem, a data-driven framework that organizes long-term conversational memory into self-configuring, low-overlap semantic views before indexing. AutoViewMem discovers candidate views from interaction traces, selects a compact complementary view set, and uses these views to guide write-time structured extraction of provenance-grounded memories. This representation-first design moves semantic disentanglement from retrieval time to write time, allowing standard top-K similarity search to retrieve focused evidence without explicit routing or iterative retrieval. We further apply offline consolidation to improve memory compactness and consistency. Experiments on the LoCoMo and PersonaMem benchmarks, under both Qwen3-8B and Qwen3-14B backbones, show that AutoViewMem improves long-horizon question answering and personalization over strong memory baselines while preserving a simple inference pipeline.

Figures & tables

Figure 1: The AutoViewMem architecture. The system processes a conversation stream through online view discovery, selects an orthogonal view set via DPP, and performs schema-guided memory extraction and retrieval to provide context for the LLM agent.

Method	Multi-Hop			Temporal			Open-Domain			Single-Hop			Overall
Method	J	F1	B1	J	F1	B1	J	F1	B1	J	F1	B1	J	F1	B1
Qwen3-8B
Full History	0.819	0.337	0.176	0.704	0.268	0.178	0.625	0.155	0.109	0.888	0.444	0.334	0.821	0.370	0.258
A-mem	0.713	0.276	0.150	0.536	0.361	0.226	0.646	0.122	0.086	0.801	0.435	0.307	0.720	0.371	0.248
MemGAS	0.752	0.157	0.081	0.741	0.107	0.036	0.740	0.104	0.042	0.815	0.229	0.125	0.783	0.183	0.093
MemoryBank	0.638	0.208	0.093	0.632	0.261	0.153	0.583	0.130	0.074	0.755	0.337	0.234	0.697	0.285	0.181

Table 1: Main Results on LoCoMo (1,540 questions) under Qwen3-8B and Qwen3-14B generation, scored by a unified Qwen3-8B judge. Each category and the overall block report LLM-Judge (J), F1, and BLEU-1 (B1). Best results among external-memory methods are in bold ; Full History is a full-context oracle (not bolded).

Method	Per-capability accuracy							Acc
Method	R-Fact	R-Ment	Track	R-Rsn	P-Rec	Gen	Sugg	Acc
Qwen3-8B
Full History	41.9	47.1	70.5	82.8	45.5	61.4	9.7	52.80
A-mem	72.1	70.6	60.4	77.8	50.9	54.4	18.3	58.06
MemGAS	65.9	58.8	62.6	68.7	43.6	29.8	8.6	50.76
MemoryBank	76.7	76.5	61.9	73.7	56.4	59.6	19.4	60.10

Table 2: Main Results on PersonaMem-32k (589 questions) under Qwen3-8B and Qwen3-14B generation. We report overall accuracy (%) under the official deterministic multiple-choice protocol and per-capability accuracy: recall of user-shared facts (R-Fact), recall of mentioned content (R-Ment), preference tracking (Track), recall of reasons (R-Rsn), preference recommendation (P-Rec), generalization (Gen), and suggestion (Sugg). Best results among external-memory methods are in bold ; Full History is a full-context oracle (not bolded).

Figure 2: Left: performance–cost trade-off on LoCoMo (x-axis: average token cost, log scale; y-axis: Judge score). Right: effect of the top-

K

retrieval budget on Judge score, comparing AutoViewMem with Mem0.

Variant	J	F1	B1	no-pos% $\downarrow$
AutoViewMem (Full)	0.837	0.456	0.339	5.14
Single-view Structured	0.774	0.422	0.317	5.53
w/o DPP Selection (random)	0.777	0.428	0.322	7.03
w/o Graph Consolidation	0.801	0.449	0.328	9.83

Table 3: Ablation on the Qwen3-8B backbone. We report overall LLM-Judge (J), F1, BLEU-1 (B1), and the no-positive rate (no-pos%, lower is better: the fraction of queries with no relevant evidence in the retrieved pool). All variants are evaluated under the same fixed

K_{0}=30

retrieval budget.

Extraction	View	Extracted memory	Information exposed beyond surface extraction
Fixed prompt	General memory	A grassy area near the office remains lit at 10 p.m. Li and others have agreed to join the user. The user does not mind differences in skill level and would like to chat afterward.	—
AutoViewMem	Football activity & venue	The user may be considering the illuminated grassy area near the office as a place to play football with Li and others after work.	Latent activity goal: links venue, time, and participants to the user’s recurring football activity, although football is not explicitly mentioned in the current utterance.
	Social motivation	For this activity, the user values participating together and socializing afterward more than differences in playing ability.	Participation priority: interprets tolerance of lower skill levels as a preference for companionship over competitive performance.
	Companionship & affect	The user expresses relief that Li and others have agreed to join and looks forward to no longer practicing alone.	Emotional significance: connects “finally” and “at last” to the user’s previously expressed frustration with solo practice.

Table 4: Case study on an ambiguous utterance. A fixed general-purpose prompt records only the surface-level content, whereas AutoViewMem’s user-adaptive views project the same evidence onto three semantically distinct aspects (— = nothing exposed beyond surface extraction). The first projection is stated tentatively because the football reading is supported by the interaction history but is not confirmed in the current utterance.

LLM Agent Memory

Large language model agent memory lets agents retain, organize, retrieve, and update information across interactions and long-running tasks.

See all 373 papers

ETAMem: Decoupling and Aligning Heterogeneous Memory for Long-Term Conversational Agents
Oct 27, 2026Xinying Qian, Ying Zhang, Yu Zhao +3Memory Management for LLM Agents Long-Term Conversational Memory
Nankai University, Tianjin, China · Tiangong University, Tianjin, China · The University of Edinburgh, Edinburgh, United Kingdom
HiGMem: A Hierarchical and LLM-Guided Memory System for Long-Term Conversational Agents
Apr 20, 2026Shuqi Cao, Jingyi He, Fei TanPersistent Memory for Language Models Hierarchical Memory
East China Normal University, Shanghai, China · Shanghai Jiao Tong University, Shanghai, China
LeanMem: Simple and Efficient Long-Term Memory for LLM Agents
Aug 4, 2026Yuxin Liao, Le Wu, Min Hou +3LLM Agent Memory Persistent Memory for Language Models
Hefei University of Technology

Oct 27, 2026J/K move · Enter open · S save

ETAMem: Decoupling and Aligning Heterogeneous Memory for Long-Term Conversational Agents

Xinying Qian, Ying Zhang, Yu Zhao, Xubo Liu+2

Nankai University, Tianjin, China · Tiangong University, Tianjin, China · The University of Edinburgh, Edinburgh, United Kingdom

Memory Management for LLM Agents Long-Term Conversational Memory Memory-Augmented Language Models Hierarchical Memory

Open paper AskPDF arXiv

Long-term memory enables Large Language Models (LLMs) to maintain coherent and personalized interactions across extended conversations. However, existing memory systems face two key challenges: (1) Information Entanglement, where heterogeneous information is mixed within coarse-grained memory units; and (2) Organizational Mismatch, where memory structures based primarily on semantic similarity may overlook other associations. To address these challenges, we propose ETAMem, a multi-view memory framework that decouples heterogeneous conversational information and aligns memory representations and organization with its distinct characteristics. Specifically, heterogeneous information parsing extracts atomic memory units and identifies affective information. View-specific memory structuring then organizes these units into three complementary hierarchical views: Event, Affect, and Time. ETAMem incrementally maintains memory structures with adaptively calibrated thresholds. For retrieval, entropy-based retrieval allocates query-specific budgets across views and performs hierarchical search within each view. Experiments on LoCoMo and LongMemEval-S demonstrate that ETAMem outperforms state-of-the-art baselines, achieving relative accuracy improvements of 12.46% and 7.96%, respectively, while maintaining competitive token efficiency. Further analyses demonstrate the complementary benefits of different memory views.

Abstract

LLM Agent Memory

Explore similar work

ETAMem: Decoupling and Aligning Heterogeneous Memory for Long-Term Conversational Agents

HiGMem: A Hierarchical and LLM-Guided Memory System for Long-Term Conversational Agents

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