CoChat

HingeMem: Boundary Guided Long-Term Memory with Query Adaptive Retrieval for Scalable Dialogues

Open access
Zhong, Yijie, Gao, Yunfan, Wang, Haofen · 2026 · Proceedings of the ACM Web Conference 2026

The DOI matches its Crossref record; 54 of the 54 references it cites were found in a registry or an index; 21 of the 27 sources its claims cite were checked for support.

Abstract

Long-term memory is critical for dialogue systems that support continuous, sustainable, and personalized interactions. However, existing methods rely on continuous summarization or OpenIE-based graph construction paired with fixed Top- k retrieval, leading to limited adaptability across query categories and high computational overhead. In this paper, we propose HingeMem, a boundary-guided long-term memory that operationalizes event segmentation theory to build an interpretable indexing interface via boundary-triggered hyperedges over four elements: person, time, location, and topic. When any such element changes, HingeMem draws a boundary and writes the current segment, thereby reducing redundant operations and preserving salient context. To enable robust and efficient retrieval under diverse information needs, HingeMem introduces query-adaptive retrieval mechanisms that jointly decide (a) what to retrieve: determine the query-conditioned routing over the element-indexed memory; (b) how much to retrieve: control the retrieval depth based on the estimated query type. Extensive experiments across LLM scales (from 0.6B to production-tier models; e.g., Qwen3-0.6B to Qwen-Flash) on LOCOMO show that HingeMem achieves approximately 20% relative improvement over strong baselines without query categories specification, while reducing computational cost (68%\downarrow question answering token cost compared to HippoRAG2). Beyond advancing memory modeling, HingeMem's adaptive retrieval makes it a strong fit for web applications requiring efficient and trustworthy memory over extended interactions.

Overview

CoChat Summary

## TL;DR

HingeMem is a neuro-inspired long-term memory system for dialogue that encodes memory at event boundaries and retrieves it with query-adaptive plans rather than fixed Top-k retrieval. On LOCOMO, it improves over strong baselines by about 5% overall, exceeds 10% on multi-hop questions, and reduces question-answering token cost by 68% versus HippoRAG2.

## Key findings

  • On LOCOMO, HingeMem achieves the best overall performance across question categories; the paper reports an over 5% improvement in overall metrics and more than 10% on multi-hop questions.
  • In the main results table, HingeMem reaches 63.9 F1, 69.4 J, and 0.406 BLEU-1 overall without category-specific question templates; with category format it reaches 75.1 J and 0.404 BLEU-1 overall.
  • The ablation table shows a progression from RAG+TM (44.6) to RAG+BM (57.4) to RAG+BM+NI (58.1), then 61.2 with hyperedge reranking and 63.9 with adaptive stop, indicating gains from both boundary-guided memory and query-adaptive retrieval.
  • The paper says HingeMem cuts question answering token cost by 68% compared to HippoRAG2 and maintains scalability across LLM sizes from 0.6B to production-tier models.

## Methods

  • The system builds boundary-guided long-term memory by detecting dialogue boundaries when key elements change across person, time, location, and topic, then storing boundary-aligned segments as hyperedges linked to element-indexed nodes.
  • It evaluates on LOCOMO, which the paper says has 10 total conversations, 27.2 average sessions per conversation, 294.1 average turns, and 15,965.8 average tokens per conversation, with 1,986 total questions across five categories.
  • Main analysis uses F1, BLEU-1, and LLM-as-a-Judge, comparing against baselines including RAG, Memorybank, LangMem, HippoRAG2, Mem0, Zep, and others under identical settings.

## Limitations

  • The paper states that existing methods can drop 30% to 40% when query categories are not given; this is a limitation of prior work, not HingeMem.
  • The text makes evident that evaluation is mainly on one benchmark, LOCOMO, so broader generalization beyond that dataset is not directly shown here.
  • The paper notes that some baseline token costs could not be measured accurately because Mem0 and Zep are official services; this is stated in the paper.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1145/3774904.3792089

Cites

54 of 54 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Verification

Claims the paper makes

11 Valid · 4 Not valid · 2 Not checked · 80 Own results and interpretations

43 more claims

Create a free account to see all 97 claims and to verify the 1 reference still open.

See all claims — free account
  • Probably not a substantial claim. 23
  • It may state two separate claims as one. 36
  • The quote only partly states this claim. 1

Create a free account to see these claims.

Citations

Cites

Research works printed but not linked: 5

  1. Institute for Basic Science. 2023. AI's memory-forming mechanism found to be strikingly similar to that of the brain. ScienceDaily (2023).
  2. Yubin Ge, Salvatore Romeo, Jason Cai, Raphael Shu, Yassine Benajiba, Monica Sunkara, and Yi Zhang. 2025. TReMu: Towards Neuro-Symbolic Temporal Reasoning for LLM-Agents with Memory in Multi-Session Dialogues. In ACL (Findings). Association for Computational Linguistics, 18974-18988.
  3. Junqing He, Liang Zhu, Rui Wang, Xi Wang, Gholamreza Haffari, and Jiaxing Zhang. 2025. MADial-Bench: Towards Real-world Evaluation of Memory-Augmented Dialogue Generation. In NAACL (Long Papers). Association for Computational Linguistics, 9902-9921.
  4. Qihong Lu, Uri Hasson, and Kenneth A Norman. 2022. A neural network model of when to retrieve and encode episodic memories. elife, Vol. 11 (2022), e74445.
  5. Jing Xu, Arthur Szlam, and Jason Weston. 2022b. Beyond Goldfish Memory: Long-Term Open-Domain Conversation. In ACL (1). Association for Computational Linguistics, 5180-5197.

Cited by

See all 1

Metadata

Authors
Zhong, Y., Gao, Y., & Wang, H.
Affiliations
Tongji University (CN)
Venue
Proceedings of the ACM Web Conference 2026
Publisher
ACM
Published
2026-04-12
Pages
3576-3587
OpenAlex
W7152670197
Cited by
1
References
8
Citations by year
2026: 1
FWCI
8.94
Type
conference paper
Language
en
Open access
Licence
cc-by
Locations
  • International World Wide Web Conference (WWW) (publishedVersion, cc-by)
  • arXiv (Cornell University) (submittedVersion, cc-by)
  • arXiv (Cornell University) (cc-by)
Funders
  • the National Natural Science Foundation of China (U23B2057)
  • Shanghai Pilot Program for Basic Research (22TQ1400300)
Topics
  • Physical Sciences › Computer Science › Artificial Intelligence › Speech and dialogue systems
  • Physical Sciences › Computer Science › Artificial Intelligence › Topic Modeling
  • Physical Sciences › Computer Science › Information Systems › Information Retrieval and Search Behavior
Keywords
long-term memory, dialogue systems, event segmentation, query routing, question answering, computational efficiency

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-10-10, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-10 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.