HingeMem: Boundary Guided Long-Term Memory with Query Adaptive Retrieval for Scalable Dialogues
Authors: Yijie Zhong, Yunfan Gao, Haofen Wang
Organizations: College of Design and Innovation, Tongji University, Shanghai, China · Shanghai Research Institute for Intelligent Autonomous Systems, Tongji University, Shanghai, China
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%↓ 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.
Figures & tables
Figure 1. Existing methods face two challenges when facing various queries: they struggle to identify key clues and find it difficult to control the retrieval process. HingeMem simulates the human cortex and hippocampus to construct boundary-guided long-term memory, employing query-adaptive retrieval to generate unique retrieval plans.
Figure 2. Overall Boundary-guided Long Term Memory construction and Query Adaptive Retrieval process of HingeMem.
Figure 3. Diagram of adaptive stop for different query types.
Single-Hop
Multi-Hop
Temporal
Open-Domain
Adversarial
Overall
Method
Cat.
F1
J
F1
J
F1
J
F1
J
F1
J
F1
J
B1
LOCOMO
✓
12.7
16.5
19.7
20.9
10.4
11.5
20.1
30.2
66.8
90.1
25.8
33.5
0.132
RAG (Top-5)
✓
32.4
47.4
26.9
38.6
22.5
18.0
21.4
35.4
90.6
90.3
42.6
50.5
0.293
RAG (Top-10)
✓
36.4
50.7
29.6
38.6
27.7
22.7
21.2
37.5
86.8
86.5
44.6
51.9
0.306
RAG (Top-20)
✓
37.7
52.5
34.9
48.9
30.0
22.7
21.9
41.6
83.2
82.9
45.5
53.5
0.315
ReadAgent †
✓
6.6
-
2.5
-
5.3
-
10.1
-
5.4
-
5.7
-
0.109
Table 1. Results on the LOCOMO dataset across five question categories. We report F1 scores ( F1 ), BLEU-1 scores ( B1 ), and LLM-as-a-Judge scores ( J ). The higher values indicate better performances. † denotes scores reported by the A-Mem paper. ‘Cat.’ indicates whether to apply the different QA format for each category. Best results are shown in Bold .
Conversation Statistics
Counts
Question Statistics
Counts
Total Conversations
10
Single-hop Questions
841
Avg. Sessions in a conversation
27.2
Multi-hop Questions
282
Avg. Turns in a conversation
294.1
Temporal Questions
321
Avg. Tokens in a conversation
15965.8
Open-domain Questions
96
Avg. Tokens in a session
587.0
Adversarial Questions
446
Avg. Tokens in a Turn
27.1
Total
1986
Table 2. Statistics of the LOCOMO dataset.
Figure 4. A comparative analysis of efficiency across different methods. For LOCOMO, we employ two configurations with context lengths of 4K and 128K. For RAG, we establish several parameter settings ranging from Top-5 to Top-20. The horizontal axis represents the total number of tokens consumed during the entire memory construction (if applicable) and the question-answering process. The circle size corresponds to the total tokens expended for all the users’ queries.
# Num
Method
All
Each Category
①
RAG+TM
44.6
36.4
29.6
27.7
21.2
86.8
②
RAG+BM
57.4
54.5
39.3
50.1
24.3
86.8
③
RAG+BM+NI
58.1
56.1
43.6
51.2
27.7
82.7
④
③+HR
61.2
59.5
47.4
52.9
28.2
86.1
⑤
③+HR+AS
63.9
61.1
53.6
57.4
30.7
87.4
Table 3. Ablation study. TM and BM represent the common Textual Memory and our Boundary Guided Memory. ② only uses the hyperedges’ description for retrieval. NI means Node Indexing, which utilizes the structural information in hyperedges. HR and AS represent the proposed Hyperedge Rerank and Adaptive Stop.
Figure 5. Analysis of the token count and its distribution of retrieval results for the question in sample ’conv-26’ in LOCOMO, comparing RAG and HingeMem. HingeMem adaptively sets the retrieval capacity for different categories of questions, thereby reducing unnecessary noise and ultimately enhancing performance and efficiency.
Figure 6. The performance comparison of different model scales in the Qwen-Series models. HingeMem delivers consistent and optimal results from the 0.6B model to the flagship model. This demonstrates the adaptability of our approach across various potential application scenarios.
Appendix figures & tables4 assets
Supplementary material from the paper’s appendix.
Appendix
Memory Statistics
Counts
Avg. Persons in a conversation
12.8
Avg. Times in a conversation
59.6
Avg. Locations in a conversation
34.6
Avg. Topics in a conversation
81.5
Avg. Hyperedges in a conversation
103.2
Appendix
Table 4. Statistics of the constructed long-term memory.
Figure 7. Performance of different values for λknee
Figure 8. Impact of the scale for precision-priority query.
Long-term conversational large language model (LLM) agents require memory systems that can recover relevant evidence from historical interactions without overwhelming the answer stage with irrelevant context. However, existing memory systems, including hierarchical ones, still often rely solely on vector similarity for retrieval. It tends to produce bloated evidence sets: adding many superficially similar dialogue turns yields little additional recall, but lowers retrieval precision, increases answer-stage context cost, and makes retrieved memories harder to inspect and manage. To address this, we propose HiGMem (Hierarchical and LLM-Guided Memory System), a two-level event-turn memory system that allows LLMs to use event summaries as semantic anchors to predict which related turns are worth reading. This allows the model to inspect high-level event summaries first and then focus on a smaller set of potentially useful turns, providing a concise and reliable evidence set through reasoning, while avoiding the retrieval overhead that would be excessively high compared to vector retrieval. On the LoCoMo10 benchmark, HiGMem achieves the best F1 on four of five question categories and improves adversarial F1 from 0.54 to 0.78 over A-Mem, while retrieving an order of magnitude fewer turns. Code is publicly available at https://github.com/ZeroLoss-Lab/HiGMem.
Shuqi Cao, Jingyi He, Fei Tan
East China Normal University, Shanghai, China · Shanghai Jiao Tong University, Shanghai, China
Large language model (LLM) agents increasingly rely on external memory to support long-horizon interaction, personalized assistance, and multi-step reasoning. However, existing memory systems still face three core challenges: they often rely too heavily on semantic similarity, which can miss evidence crucial for user-centric understanding; they frequently store related experiences as isolated fragments, weakening temporal and causal coherence; and they typically use static memory granularities that do not adapt well to the requirements of different questions. We propose AdaMem, an adaptive user-centric memory framework for long-horizon dialogue agents. AdaMem organizes dialogue history into working, episodic, persona, and graph memories, enabling the system to preserve recent context, structured long-term experiences, stable user traits, and relation-aware connections within a unified framework. At inference time, AdaMem first resolves the target participant, then builds a question-conditioned retrieval route that combines semantic retrieval with relation-aware graph expansion only when needed, and finally produces the answer through a role-specialized pipeline for evidence synthesis and response generation. We evaluate AdaMem on the LoCoMo and PERSONAMEM benchmarks for long-horizon reasoning and user modeling. Experimental results show that AdaMem achieves state-of-the-art performance on both benchmarks. The code will be released upon acceptance.
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.