Elamur Research⌕
☰
home / papers / 2026 / Preference-Aware Memory Update for Long-Term LLM Agents

Preference-Aware Memory Update for Long-Term LLM Agents

Paper · 2026 · Haoran Sun, Zekun Zhang and Shaoning Zeng

Summary

Long-running LLM agents usually rely on memory systems such as MemoryBank, MemGPT, MemInsight or A-MEM. These systems mostly improve how memory is stored and retrieved. They pay little attention to keeping memory current as a user's preferences change over time. The authors propose PAMU (Preference-Aware Memory Update), a plug-in module that needs no fine-tuning and works with any underlying model. After each turn, PAMU builds a five-dimensional preference profile from the dialogue using off-the-shelf components: a RoBERTa tone classifier, an emotion classifier, an OpenIE-based information-density measure, a formality classifier and a normalized response-length score. For each dimension, it tracks a short-horizon sliding-window average and a long-horizon EMA and blends the two with a weight λ. When the gap between them crosses a threshold δ, PAMU treats this as a preference shift and rewrites the prompt and memory. Tested on five task scenarios of LoCoMo, PAMU improved output quality for five baseline memory systems. A case study also shows that it adapts in real time where the baselines lag. The work matters because it gives agents an interpretable, low-cost way to follow changes in what users want.

Key points

  • Preference dimensions: PAMU tracks five dimensions:
    • tone style and emotional tone, each stored as a category plus its probability;
    • response length, information density (OpenIE triples per word) and formality, each a normalized scalar.
  • Update rule: The fused preference estimate is λ·SW + (1−λ)·EMA. A memory update fires only when the difference |SW − EMA| exceeds δ.
  • Hyperparameter stability: The paper reports stable performance for λ ∈ [0.4, 0.7] and EMA decay β ∈ [0.8, 0.95]. Ablations were run on temporal reasoning with LLaMA 7B, scored by the average of F1 and BLEU-1.
  • Case study: In a simulated dialogue where the user's preference shifts, both GPT-4 and ten human annotators judged the responses. Alignment scores (1–5, GPT/human) rose from 2.1/2.2 without PAMU to 4.8/4.5 with it. Without PAMU, the model adapted with a two-turn delay; with PAMU, it adapted immediately.
  • Limitations stated by the authors:
    • errors in the external classifiers carry over into the preference profile;
    • prompt-based control can be brittle;
    • converting scores into discrete labels loses detail;
    • the preference set is narrow and leaves out, for example, reasoning style and persona.

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

 Institutions

🎓University of Electronic Science and Technology of ChinaChengdu · China31 works · 2,012 citations

 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

Cite

Cite as Haoran Sun, Zekun Zhang and Shaoning Zeng. 2026. Preference-Aware Memory Update for Long-Term 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.38

@inproceedings{sun2026preference,
  title = {{Preference-Aware Memory Update for Long-Term LLM Agents}},
  author = {Haoran Sun, Zekun Zhang and Shaoning Zeng},
  booktitle = {Findings of the Association for Computational Linguistics: ACL 2026},
  year = {2026},
  doi = {10.18653/v1/2026.findings-acl.38},
  url = {https://doi.org/10.18653/v1/2026.findings-acl.38}
}

References

25 references, 10 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. DeepSeek-AI. 2025. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. Preprint, arXiv:2501.12948.
  2. Jiawei Gu, Xuhui Jiang, Zhichao Shi, Hexiang Tan, Xuehao Zhai, Chengjin Xu, Wei Li, Yinghan Shen, Shengjie Ma, Honghao Liu, and 1 others. 2024. A survey on llm-as-a-judge. arXiv preprint arXiv:2411.15594. find
  3. Taicheng Guo, Xiuying Chen, Yaqi Wang, Ruidi Chang, Shichao Pei, Nitesh V Chawla, Olaf Wiest, and Xiangliang Zhang. 2024. Large language model based multi-agents: A survey of progress and challenges. arXiv preprint arXiv:2402.01680.
  4. Xu Han, Tianyu Gao, Yuan Yao, Demin Ye, Zhiyuan Liu, and Maosong Sun. 2019. Opennre: An open and extensible toolkit for neural relation extraction. arXiv preprint arXiv:1909.13078. find
  5. Xu Huang, Weiwen Liu, Xiaolong Chen, Xingmei Wang, Hao Wang, Defu Lian, Yasheng Wang, Ruiming Tang, and Enhong Chen. 2024a. Understanding the planning of llm agents: A survey. arXiv preprint arXiv:2402.02716. find
  6. Xu Huang, Weiwen Liu, Xiaolong Chen, Xingmei Wang, Hao Wang, Defu Lian, Yasheng Wang, Ruiming Tang, and Enhong Chen. 2024b. Understanding the planning of llm agents: A survey. arXiv preprint arXiv:2402.02716. find
  7. Haolin Jin, Linghan Huang, Haipeng Cai, Jun Yan, Bo Li, and Huaming Chen. 2024. From llms to llm-based agents for software engineering: A survey of current, challenges and future. arXiv preprint arXiv:2408.02479. — [arXiv:2408.02479](https://arxiv.org/abs/2408.02479) find
  8. Kuang-Huei Lee, Xinyun Chen, Hiroki Furuta, John Canny, and Ian Fischer. 2024. A human-inspired reading agent with gist memory of very long contexts. Preprint, arXiv:2402.09727.
  9. Yuanchun Li, Hao Wen, Weijun Wang, Xiangyu Li, Yizhen Yuan, Guohong Liu, Jiacheng Liu, Wenxing Xu, Xiang Wang, Yi Sun, and 1 others. 2024. Personal llm agents: Insights and survey about the capability, efficiency and security. arXiv preprint arXiv:2401.05459. find
  10. Adyasha Maharana, Dong-Ho Lee, Sergey Tulyakov, Mohit Bansal, Francesco Barbieri, and Yuwei Fang. 2024. Evaluating very long-term conversational memory of llm agents. arXiv preprint arXiv:2402.17753. find
  11. Charles Packer, Vivian Fang, Shishir_G Patil, Kevin Lin, Sarah Wooders, and Joseph_E Gonzalez. 2023. Memgpt: Towards llms as operating systems. find
  12. Rana Salama, Jason Cai, Michelle Yuan, Anna Currey, Monica Sunkara, Yi Zhang, and Yassine Benajiba. 2025. Meminsight: Autonomous memory augmentation for llm agents. arXiv preprint arXiv:2503.21760.
  13. Haoran Sun, Haoyu Bian, Shaoning Zeng, Yunbo Rao, Xu Xu, Lin Mei, and Jianping Gou. 2025. Datasetagent: A novel multi-agent system for auto-constructing datasets from real-world images. Preprint, arXiv:2507.08648. — [arXiv:2507.08648](https://arxiv.org/abs/2507.08648) find
  14. Haoran Sun and Shaoning Zeng. 2025. Introspection of thought helps ai agents. Preprint, arXiv:2507.08664. — [arXiv:2507.08664](https://arxiv.org/abs/2507.08664) find
  15. Haoran Sun, Shaoning Zeng, and Bob Zhang. 2026. H-MEM: Hierarchical memory for high-efficiency long-term reasoning in LLM agents. In Proceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers), pages 341–350, Rabat, Morocco. Ass…
  16. Hao Tian, Can Gao, Xinyan Xiao, Hao Liu, Bolei He, Hua Wu, Haifeng Wang, and Feng Wu. 2020. Skep: Sentiment knowledge enhanced pre-training for sentiment analysis. Preprint, arXiv:2005.05635.
  17. Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. 2023. Llama: Open and efficient foundation language models. Preprint…
  18. Bing Wang, Xinnian Liang, Jian Yang, Hui Huang, Shuangzhi Wu, Peihao Wu, Lu Lu, Zejun Ma, and Zhoujun Li. 2023. Enhancing large language model with self-controlled memory framework. arXiv preprint arXiv:2304.13343. — [arXiv:2304.13343](https://arxiv.org/abs/2304.13343) find
  19. Junchao Wu, Shu Yang, Runzhe Zhan, Yulin Yuan, Lidia Sam Chao, and Derek Fai Wong. 2025. A survey on llm-generated text detection: Necessity, methods, and future directions. Computational Linguistics, pages 1–66.
  20. Wujiang Xu, Zujie Liang, Kai Mei, Hang Gao, Juntao Tan, and Yongfeng Zhang. 2025. A-mem: Agentic memory for llm agents. arXiv preprint arXiv:2502.12110.
  21. An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, Guanting Dong, Haoran Wei, Huan Lin, Jialong Tang, Jialin Wang, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Ma, and 40 others. 2024. Qwen2 technical report. arXiv preprint arXi… find
  22. Yifan Yao, Jinhao Duan, Kaidi Xu, Yuanfang Cai, Zhibo Sun, and Yue Zhang. 2024. A survey on large language model (llm) security and privacy: The good, the bad, and the ugly. High-Confidence Computing, page 100211.
  23. Zeyu Zhang, Xiaohe Bo, Chen Ma, Rui Li, Xu Chen, Quanyu Dai, Jieming Zhu, Zhenhua Dong, and Ji-Rong Wen. 2024a. A survey on the memory mechanism of large language model based agents. arXiv preprint arXiv:2404.13501. find
  24. Zeyu Zhang, Xiaohe Bo, Chen Ma, Rui Li, Xu Chen, Quanyu Dai, Jieming Zhu, Zhenhua Dong, and Ji-Rong Wen. 2024b. A survey on the memory mechanism of large language model based agents. arXiv preprint arXiv:2404.13501. find
  25. Wanjun Zhong, Lianghong Guo, Qiqi Gao, He Ye, and Yanlin Wang. 2024. Memorybank: Enhancing large language models with long-term memory. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 19724–19731. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/sun-2026-preference-aware-memory-update-long/