Elamur Research⌕
☰
home / papers / 2026 / Beyond Dialogue Time: Temporal Semantic Memory for Personalized LLM Agents

Beyond Dialogue Time: Temporal Semantic Memory for Personalized LLM Agents

Paper · 2026 · Miao Su, Yucan Guo, Zhongni Hou et al.

Summary

Memory modules let LLM agents personalize answers across long, multi-session conversations, but current systems handle time poorly. They index facts by the moment a message was sent, not by when the described event actually occurs. They also store isolated point-wise records, so an ongoing state such as a week-long trip gets split into unrelated fragments.

The authors propose Temporal Semantic Memory (TSM). During construction, it extracts episodic facts into a temporal knowledge graph tied to event time, following Zep's approach. It then merges temporally adjacent, semantically related entries into durative memories, which are topic and persona summaries grouped by month. At query time, TSM infers the question's temporal intent and combines dense retrieval with temporal reranking and filtering. This supplies the answer generator with context that is valid for the right time period. Incremental online updates and periodic consolidation keep the memory current.

On LongMemEval and LoCoMo, TSM outperforms memory baselines by up to 12.2 absolute accuracy points. The largest gains appear on multi-session and temporal-reasoning questions. This shows that modeling event time and duration matters for reliable agent memory.

Key points

  • Two problems are targeted:
    • Temporal inaccuracy: dialogue time is confused with event time.
    • Temporal fragmentation: only point-wise memories are kept.
  • TSM has three parts:
    • episodic memory, stored as a temporal knowledge graph;
    • durative memory, made of topic and persona summaries at fixed monthly granularity;
    • semantic-time-guided retrieval.
  • On LongMemEval_S, TSM reaches 74.80% accuracy, compared with A-MEM 62.60, Mem0 53.61 and MemoryOS 44.80. It also has the lowest recall latency, at 1.57 s (P50) and 2.39 s (P95).
  • The ablation on LongMemEval_S shows both components help:
    • Removing temporal reranking and filtering lowers accuracy to 72.8 (−2.0), with a −6.0 drop on temporal questions.
    • Removing summaries lowers accuracy to 73.6 and costs −10.0 on single-session preference questions.
  • A sleep-time (offline) update variant cuts total token cost from 2,065.54k to 1,959.53k without affecting online latency. The authors list the fixed monthly granularity as a limitation.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Fields, topics and concepts

📚 Shelves 4The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksArts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 works

🧭 Topics 1Research subjects and languages the work is about.Dialogue systems · 624 works

Cite

Cite as Miao Su, Yucan Guo, Zhongni Hou, Long Bai, Zixuan Li, Yufei Zhang, Guojun Yin, Wei Lin, Xiaolong Jin, Jiafeng Guo and Xueqi Cheng. 2026. Beyond Dialogue Time: Temporal Semantic Memory for Personalized LLM Agents. Findings of the Association for Computational Linguistics: ACL 2026.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.findings-acl.1496

@inproceedings{su2026beyond,
  title = {{Beyond Dialogue Time: Temporal Semantic Memory for Personalized LLM Agents}},
  author = {Miao Su, Yucan Guo, Zhongni Hou, Long Bai, Zixuan Li, Yufei Zhang, Guojun Yin, Wei Lin, Xiaolong Jin, Jiafeng Guo and Xueqi Cheng},
  booktitle = {Findings of the Association for Computational Linguistics: ACL 2026},
  year = {2026},
  doi = {10.18653/v1/2026.findings-acl.1496},
  url = {https://doi.org/10.18653/v1/2026.findings-acl.1496}
}

References

19 references, 5 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. ByteDance. 2025. Deerflow: Deep exploration and efficient research framework. https://deerflow.tech/z. ↗
  2. Camel-AI. 2025. Owl: Optimized workforce learning for general multi-agent assistance in real-world task automation. find
  3. Prateek Chhikara, Dev Khant, Saket Aryan, Taranjeet Singh, and Deshraj Yadav. 2025a. Mem0: Building production-ready ai agents with scalable long-term memory. arXiv preprint arXiv:2504.19413.
  4. Prateek Chhikara, Dev Khant, Saket Aryan, Taranjeet Singh, and Deshraj Yadav. 2025b. Mem0: Building Production-Ready AI Agents with Scalable Long-Term Memory.
  5. Jizhan Fang, Xinle Deng, Haoming Xu, Ziyan Jiang, Yuqi Tang, Ziwen Xu, Shumin Deng, Yunzhi Yao, Mengru Wang, Shuofei Qiao, Huajun Chen, and Ningyu Zhang. 2025. LightMem: Lightweight and Efficient Memory-Augmented Generation. arXiv preprint. ArXiv:2510.18866 [cs] TLDR: Inspired by the Atkinson-Shiffr… — [arXiv:2510.18866](https://arxiv.org/abs/2510.18866) find
  6. Google. 2025. Gemini deep research — your personal research assistant. https://gemini.google/ overview/deep-research/?hl=en-GB. ↗
  7. Matthew Honnibal, Ines Montani, Sofie Van Lan-deghem, and Adriane Boyd. 2020. spacy: Industrial-strength natural language processing in python. find
  8. Yuyang Hu, Shichun Liu, Yanwei Yue, Guibin Zhang, Boyang Liu, Fangyi Zhu, Jiahang Lin, Honglin Guo, Shihan Dou, Zhiheng Xi, Senjie Jin, Jiejun Tan, Yanbin Yin, Jiongnan Liu, Zeyu Zhang, Zhongxiang Sun, Yutao Zhu, Hao Sun, Boci Peng, and 28 others. 2025. Memory in the Age of AI Agents. arXiv preprint… — [arXiv:2512.13564](https://arxiv.org/abs/2512.13564) find
  9. Haoyu Huang, Yongfeng Huang, Junjie Yang, Zhenyu Pan, Yongqiang Chen, Kaili Ma, Hongzhi Chen, and James Cheng. 2025. Retrieval-Augmented Generation with Hierarchical Knowledge. arXiv preprint. ArXiv:2503.10150 [cs] TLDR: This paper introduces a new RAG approach, called HiRAG, which utilizes hierarch… find
  10. Alexis Huet, Zied Ben Houidi, and Dario Rossi. 2025. Episodic Memories Generation and Evaluation Benchmark for Large Language Models. arXiv preprint. ArXiv:2501.13121 [cs] TLDR: It is argued that integrating episodic memory capabilities into LLM is essential for advancing AI towards human-like cogni… — [arXiv:2501.13121](https://arxiv.org/abs/2501.13121) find
  11. David Kadavy. 2021. Digital Zettelkasten: Principles, Methods, & Examples. Kadavy, Inc. Google-Books-ID: o4gwEAAAQBAJ. find
  12. Jiazheng Kang, Mingming Ji, Zhe Zhao, and Ting Bai. Memory OS of AI Agent.
  13. Sangyeop Kim, Yohan Lee, Sanghwa Kim, Hyunjong Kim, and Sungzoon Cho. 2025. Pre-Storage Reasoning for Episodic Memory: Shifting Inference Burden to Memory for Personalized Dialogue. arXiv preprint. ArXiv:2509.10852 [cs] TLDR: This work introduces PREMem (Pre-storage Reasoning for Episodic Memory), a… find
  14. Taeyoon Kwon, Dongwook Choi, Sunghwan Kim, Hyojun Kim, Seungjun Moon, Beong-woo Kwak, Kuan-Hao Huang, and Jinyoung Yeo. 2025. Embodied Agents Meet Personalization: Exploring Memory Utilization for Personalized Assistance. arXiv preprint. ArXiv:2505.16348 [cs] TLDR: ME-MENTO is presented, a personali… — [arXiv:2505.16348](https://arxiv.org/abs/2505.16348) find
  15. Hao Li, Chenghao Yang, An Zhang, Yang Deng, Xiang Wang, and Tat-Seng Chua. 2025. Hello Again! LLM-powered Personalized Agent for Long-term Dialogue. arXiv preprint. ArXiv:2406.05925 [cs].
  16. Xinbin Liang, Jinyu Xiang, Zhaoyang Yu, Jiayi Zhang, Sirui Hong, Sheng Fan, and Xiao Tang. 2025. Openmanus: An open-source framework for building general ai agents. find
  17. Junyu Luo, Weizhi Zhang, Ye Yuan, Yusheng Zhao, Junwei Yang, Yiyang Gu, Bohan Wu, Binqi Chen, Ziyue Qiao, Qingqing Long, Rongcheng Tu, Xiao Luo, Wei Ju, Zhiping Xiao, Yifan Wang, Meng Xiao, Chenwu Liu, Jingyang Yuan, Shichang Zhang, and 7 others. 2025. Large Language Model Agent: A Survey on Methodo…
  18. Christopher J. MacDonald, Kyle Q. Lepage, Uri T. Eden, and Howard Eichenbaum. 2011. Hippocampal "time cells" bridge the gap in memory for discontiguous events. Neuron, 71(4):737–749. TLDR: A robust hippocampal representation of sequence memories is reported, highlighted by "time cells" that encode s… find
  19. Adyasha Maharana, Dong-Ho Lee, Sergey Tulyakov, Mohit Bansal, Francesco Barbieri, and Yuwei Fang. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/su-2026-beyond-dialogue-time-temporal-semantic/