---
title: "MetaMem — paper by Xin et al. (2026) — summary and ideas"
description: "External memory systems allow LLMs to handle long-term user interactions that exceed their context window. 25 references."
authors: ["Haidong Xin", "Xinze Li", "Zhenghao Liu", "Yukun Yan", "Shuo Wang", "Cheng Yang", "Yu Gu", "Ge Yu", "Maosong Sun"]
year: 2026
venue: "Findings of the Association for Computational Linguistics: ACL 2026"
source: "https://aclanthology.org/2026.findings-acl.270/"
doi: "10.18653/v1/2026.findings-acl.270"
licence: "CC BY 4.0 (ACL Anthology)"
licence_url: "https://creativecommons.org/licenses/by/4.0/"
summary_licence: "Elamur Cosmic License"
canonical: "https://research.elamur.ai/free/en/papers/xin-2026-metamem-evolving-meta-memory-knowledge/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/xin-2026-metamem-evolving-meta-memory-knowledge/)

[home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2026](https://md.research.elamur.ai/free/en/papers/2026/page/41/) / MetaMem: Evolving Meta-Memory for Knowledge Utilization through Self-Reflective Symbolic Optimization

# MetaMem: Evolving Meta-Memory for Knowledge Utilization through Self-Reflective Symbolic Optimization

**Authors:** Haidong Xin, [Xinze Li](https://md.research.elamur.ai/free/en/authors/xinze-li/), [Zhenghao Liu](https://md.research.elamur.ai/free/en/authors/zhenghao-liu/), [Yukun Yan](https://md.research.elamur.ai/free/en/authors/yukun-yan/), [Shuo Wang](https://md.research.elamur.ai/free/en/authors/shuo-wang/), [Cheng Yang](https://md.research.elamur.ai/free/en/authors/cheng-hong-yang/), [Yu Gu](https://md.research.elamur.ai/free/en/authors/yu-gu/), [Ge Yu](https://md.research.elamur.ai/free/en/authors/ge-yu/) and [Maosong Sun](https://md.research.elamur.ai/free/en/authors/maosong-sun/) · **Year:** 2026 · **Venue:** [Findings of the Association for Computational Linguistics: ACL 2026](https://md.research.elamur.ai/free/en/venues/findings/) ([Findings of the Association for Computational Linguistics 2026](https://md.research.elamur.ai/free/en/venues/findings/2026/)) · **Publisher:** [Association for Computational Linguistics](https://md.research.elamur.ai/free/en/publishers/association-for-computational-linguistics/) · **Licence:** [CC BY 4.0 (ACL Anthology)](https://creativecommons.org/licenses/by/4.0/)

**Links:** [Publisher's page](https://aclanthology.org/2026.findings-acl.270/) · [PDF](https://aclanthology.org/2026.findings-acl.270.pdf) · [DOI 10.18653/v1/2026.findings-acl.270](https://doi.org/10.18653/v1/2026.findings-acl.270) · [OpenAlex](https://openalex.org/W7166851116)

**Citations:** 0 citations (OpenAlex, 2026-10-14)

## Summary
External memory systems allow LLMs to handle long-term user interactions that exceed their context window. However, current ways of building memory, such as topic clustering or summarization, often split sessions into pieces that lose their temporal and logical links. Retrieved fragments can then overlap or conflict, for example when a trip segment's mileage is counted alongside the whole trip's total, and the model reasons poorly. MetaMem does not change how memory is built. Instead, it adds a learned "meta-memory": a set of explicit, reusable rules for how to use retrieved fragments. Starting from an empty set, the meta-memory is optimized over mini-batches of multi-task data. The LLM generates several reasoning attempts, reflects on them using correctness-based rewards, and edits the experience set. At inference time, these rules are added to the prompt. The approach outperforms strong baselines by more than 3.6%, generalizes across memory systems and from out-of-domain data, and improves as training data grows. This suggests memory systems could keep improving through continued use.

## Key points
- The meta-memory is a set of textual experience units updated iteratively through self-reflection. No model weights change, and the method works on top of a LightMem-style topic-aggregated memory.
- The reported gain over all baselines is above 3.6%. Code and data are released at github.com/OpenBMB/MetaMem.
- During training, the share of task-agnostic ("General") units grew from about 65% to over 80%. Most tasks peaked around step 5 and then declined because of overfitting, with the Knowledge Update task as the exception.
- In scaling tests with Qwen3-30B-A3B-Instruct trained on ShareGPT, MetaMem overtook the baseline at 400 samples and led by 2.9% at 1,000. Output perplexity fell from 3.59 to 3.19, with little further change after 800 samples.
- Limitations: the method depends on an LLM-based judge, and out-of-domain meta-memory still trails in-domain training.

*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

- [Tsinghua University](https://md.research.elamur.ai/free/en/universities/tsinghua-university/) — University, Beijing, China · 415 works · 94,028 citations
- [Beijing University of Posts and Telecommunications](https://md.research.elamur.ai/free/en/universities/beijing-university-of-posts-and-telecommunications/) — University, Beijing, China · 78 works · 33,365 citations
- [Northeastern University](https://md.research.elamur.ai/free/en/universities/northeastern-university-cn/) — University, Shenyang, China · 53 works · 2,205 citations

📚 Shelves (4): [🖥️ Computer Science](https://md.research.elamur.ai/free/en/shelves/computer-science/) · 18,614 works · [🤖 Artificial Intelligence](https://md.research.elamur.ai/free/en/shelves/computer-science/artificial-intelligence/) · 17,965 works · [🎭 Arts and Humanities](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/) · 17,450 works · [🗣️ Language and Linguistics](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/language-and-linguistics/) · 17,296 works

## Cite

Cite as: Haidong Xin, Xinze Li, Zhenghao Liu, Yukun Yan, Shuo Wang, Cheng Yang, Yu Gu, Ge Yu and Maosong Sun. 2026. MetaMem: Evolving Meta-Memory for Knowledge Utilization through Self-Reflective Symbolic Optimization. Findings of the Association for Computational Linguistics: ACL 2026.

The paper itself, not this page. BibTeX from the library's record of the work.

```bibtex
@inproceedings{xin2026metamem,
  title = {{MetaMem: Evolving Meta-Memory for Knowledge Utilization through Self-Reflective Symbolic Optimization}},
  author = {Haidong Xin, Xinze Li, Zhenghao Liu, Yukun Yan, Shuo Wang, Cheng Yang, Yu Gu, Ge Yu and Maosong Sun},
  booktitle = {Findings of the Association for Computational Linguistics: ACL 2026},
  year = {2026},
  doi = {10.18653/v1/2026.findings-acl.270},
  url = {https://doi.org/10.18653/v1/2026.findings-acl.270}
}
```

- [← previous Metalinguistic negation in Jordanian Arabic: Clause-initial <em>miš/mū </em>as a high objection operator](https://md.research.elamur.ai/free/en/papers/khalaf-2026-metalinguistic-negation-jordanian-arabic-clause/)
- [next → MetaMiners at SMM4H-HeaRD 2026: A Semantic-Structural Knowledge-Enriched Ensemble for SARS-CoV-2 Metadata Identification](https://md.research.elamur.ai/free/en/papers/ursu-2026-metaminers-smm4h-heard-2026-semantic/)
- [↑ back to papers of 2026](https://md.research.elamur.ai/free/en/papers/2026/page/41/)

## References

25 references, 14 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

1. Prateek Chhikara, Dev Khant, Saket Aryan, Taranjeet Singh, and Deshraj Yadav. 2025. [Mem0: Building production-ready ai agents with scalable long-term memory](https://md.research.elamur.ai/free/en/papers/chhikara-2025-mem0-building-production-ready-ai/). ArXiv preprint, abs/2504.19413.
2. Jizhan Fang, Xinle Deng, Haoming Xu, Ziyan Jiang, Yuqi Tang, Ziwen Xu, Shumin Deng, Yunzhi Yao, Mengru Wang, Shuofei Qiao, and 1 others. 2025. Lightmem: Lightweight and efficient memory-augmented generation. ArXiv preprint, abs/2510.18866. — [arXiv:2510.18866](https://arxiv.org/abs/2510.18866) [find](https://scholar.google.com/scholar?q=Jizhan+Fang%2C+Xinle+Deng%2C+Haoming+Xu%2C+Ziyan+Jiang%2C+Yuqi+Tang%2C+Ziwen+Xu%2C+Shumin+Deng%2C+Yunzhi+Yao%2C+Mengru+Wang%2C+Shuofei+Qiao%2C+and+1+others%2E+2025%2E+Lightmem%3A+Lightweight+and+efficient+memory%2Daugmented+gene)
3. John H Flavell. 1979. Metacognition and cognitive monitoring: A new area of cognitive–developmental inquiry. American psychologist, 34(10):906. [find](https://scholar.google.com/scholar?q=John+H+Flavell%2E+1979%2E+Metacognition+and+cognitive+monitoring%3A+A+new+area+of+cognitive%E2%80%93developmental+inquiry%2E+American+psychologist%2C+34%2810%29%3A906%2E)
4. Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, and 1 others. 2024. [The llama 3 herd of models](https://md.research.elamur.ai/free/en/papers/grattafiori-2024-llama-3-herd-models/). ArXiv preprint, abs/2407.21783.
5. Junxian He, Graham Neubig, and Taylor Berg-Kirkpatrick. 2021. [Efficient nearest neighbor language models](https://md.research.elamur.ai/free/en/papers/he-2021-efficient-nearest-neighbor-language-models/). In Proceedings of EMNLP, pages 5703–5714.
6. Jiazheng Kang, Mingming Ji, Zhe Zhao, and Ting Bai. 2025. [Memory os of ai agent](https://md.research.elamur.ai/free/en/papers/kang-2025-memory-os-ai-agent/). ArXiv preprint, abs/2506.06326.
7. Patrick S. H. Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, Sebastian Riedel, and Douwe Kiela. 2020. [Retrieval-augmented generation for knowledge-intensive NLP tasks](https://md.research.elamur.ai/free/en/papers/lewis-2020-retrieval-augmented-generation-knowledge-intensive/). In Proceedings of NeurIPS.
8. Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, and 1 others. 2024a. Deepseek-v3 technical report. ArXiv preprint, abs/2412.19437. [find](https://scholar.google.com/scholar?q=Aixin+Liu%2C+Bei+Feng%2C+Bing+Xue%2C+Bingxuan+Wang%2C+Bochao+Wu%2C+Chengda+Lu%2C+Chenggang+Zhao%2C+Chengqi+Deng%2C+Chenyu+Zhang%2C+Chong+Ruan%2C+and+1+others%2E+2024a%2E+Deepseek%2Dv3+technical+report%2E+ArXiv+preprint%2C+abs%2F2412)
9. Jiaheng Liu, Dawei Zhu, Zhiqi Bai, Yancheng He, Huanxuan Liao, Haoran Que, Zekun Wang, Chenchen Zhang, Ge Zhang, Jiebin Zhang, and 1 others. 2025. [A comprehensive survey on long context language modeling](https://md.research.elamur.ai/free/en/papers/liu-2025-comprehensive-survey-long-context-language/). arXiv preprint arXiv:2503.17407.
10. Nelson F. Liu, Kevin Lin, John Hewitt, Ashwin Paranjape, Michele Bevilacqua, Fabio Petroni, and Percy Liang. 2024b. [Lost in the middle: How language models use long contexts](https://md.research.elamur.ai/free/en/papers/liu-2023-lost-middle-language-models-use/). TACL, 12:157–173.
11. Zhenghao Liu, Pengcheng Huang, Zhipeng Xu, Xinze Li, Shuliang Liu, Chunyi Peng, Haidong Xin, Yukun Yan, Shuo Wang, Xu Han, and 1 others. 2026. Knowledge intensive agents. AI Open. [find](https://scholar.google.com/scholar?q=Zhenghao+Liu%2C+Pengcheng+Huang%2C+Zhipeng+Xu%2C+Xinze+Li%2C+Shuliang+Liu%2C+Chunyi+Peng%2C+Haidong+Xin%2C+Yukun+Yan%2C+Shuo+Wang%2C+Xu+Han%2C+and+1+others%2E+2026%2E+Knowledge+intensive+agents%2E+AI+Open%2E)
12. Charles Packer, Vivian Fang, Shishir G. Patil, Kevin Lin, Sarah Wooders, and Joseph E. Gonzalez. 2023. Memgpt: Towards llms as operating systems. ArXiv preprint, abs/2310.08560. [find](https://scholar.google.com/scholar?q=Charles+Packer%2C+Vivian+Fang%2C+Shishir+G%2E+Patil%2C+Kevin+Lin%2C+Sarah+Wooders%2C+and+Joseph+E%2E+Gonzalez%2E+2023%2E+Memgpt%3A+Towards+llms+as+operating+systems%2E+ArXiv+preprint%2C+abs%2F2310%2E08560%2E)
13. Joon Sung Park, Joseph O’Brien, Carrie Jun Cai, Meredith Ringel Morris, Percy Liang, and Michael S Bernstein. 2023. Generative agents: Interactive simulacra of human behavior. In Proceedings of UIST, pages 1–22. [find](https://scholar.google.com/scholar?q=Joon+Sung+Park%2C+Joseph+O%E2%80%99Brien%2C+Carrie+Jun+Cai%2C+Meredith+Ringel+Morris%2C+Percy+Liang%2C+and+Michael+S+Bernstein%2E+2023%2E+Generative+agents%3A+Interactive+simulacra+of+human+behavior%2E+In+Proceedings+of+UIST%2C+)
14. ShareAI Lab. 2023. Sharegpt-chinese-english-90k: A bilingual chinese-english human-machine dialogue dataset. [https://huggingface.co/datasets/](https://huggingface.co/datasets/) shareAI/ShareGPT-Chinese-English-90k. Hugging Face dataset repository. [↗](https://huggingface.co/datasets/)
15. Aohan Zeng, Xin Lv, Qinkai Zheng, Zhenyu Hou, Bin Chen, Chengxing Xie, Cunxiang Wang, Da Yin, Hao Zeng, Jiajie Zhang, and 1 others. 2025. [Glm-4.5: Agentic, reasoning, and coding (arc) foundation models](https://md.research.elamur.ai/free/en/papers/team-2025-4-glm-4-5-agentic/). ArXiv preprint, abs/2508.06471.
16. Guibin Zhang, Muxin Fu, Guancheng Wan, Miao Yu, Kun Wang, and Shuicheng Yan. 2025. G-memory: Tracing hierarchical memory for multi-agent systems. ArXiv preprint, abs/2506.07398. — [arXiv:2506.07398](https://arxiv.org/abs/2506.07398) [find](https://scholar.google.com/scholar?q=Guibin+Zhang%2C+Muxin+Fu%2C+Guancheng+Wan%2C+Miao+Yu%2C+Kun+Wang%2C+and+Shuicheng+Yan%2E+2025%2E+G%2Dmemory%3A+Tracing+hierarchical+memory+for+multi%2Dagent+systems%2E+ArXiv+preprint%2C+abs%2F2506%2E07398%2E+%E2%80%94+%5BarXiv%3A2506%2E07398%5D%28h)
17. Wanjun Zhong, Lianghong Guo, Qiqi Gao, He Ye, and Yanlin Wang. 2024. Memorybank: Enhancing large language models with long-term memory. In Proceedings of AAAI, pages 19724–19731. [find](https://scholar.google.com/scholar?q=Wanjun+Zhong%2C+Lianghong+Guo%2C+Qiqi+Gao%2C+He+Ye%2C+and+Yanlin+Wang%2E+2024%2E+Memorybank%3A+Enhancing+large+language+models+with+long%2Dterm+memory%2E+In+Proceedings+of+AAAI%2C+pages+19724%E2%80%9319731%2E)
18. Zihan Zhou, Chong Li, Xinyi Chen, Shuo Wang, Yu Chao, Zhili Li, Haoyu Wang, Qi Shi, Zhixing Tan, Xu Han, and 1 others. 2025. Llm× mapreduce: Simplified long-sequence processing using large language models. In Proceedings of ACL, pages 27664– 27678. [find](https://scholar.google.com/scholar?q=Zihan+Zhou%2C+Chong+Li%2C+Xinyi+Chen%2C+Shuo+Wang%2C+Yu+Chao%2C+Zhili+Li%2C+Haoyu+Wang%2C+Qi+Shi%2C+Zhixing+Tan%2C+Xu+Han%2C+and+1+others%2E+2025%2E+Llm%C3%97+mapreduce%3A+Simplified+long%2Dsequence+processing+using+large+language+mo)
19. Kimi Team, Yifan Bai, Yiping Bao, Guanduo Chen, Jiahao Chen, Ningxin Chen, Ruijue Chen, Yanru Chen, Yuankun Chen, Yutian Chen, and 1 others. 2025. [Kimi k2: Open agentic intelligence](https://md.research.elamur.ai/free/en/papers/kimi-team-2025-kimi-k2-open-agentic/). ArXiv preprint, abs/2507.20534.
20. Harsh Trivedi, Niranjan Balasubramanian, Tushar Khot, and Ashish Sabharwal. 2023. [Interleaving retrieval with chain-of-thought reasoning for knowledge-intensive multi-step questions](https://md.research.elamur.ai/free/en/papers/trivedi-2022-interleaving-retrieval-chain-thought-reasoning/). In Proceedings of ACL, pages 10014–10037.
21. Wenhui Wang, Furu Wei, Li Dong, Hangbo Bao, Nan Yang, and Ming Zhou. 2020. Minilm: Deep self-attention distillation for task-agnostic compression of pre-trained transformers. In Proceedings of NeurIPS. [find](https://scholar.google.com/scholar?q=Wenhui+Wang%2C+Furu+Wei%2C+Li+Dong%2C+Hangbo+Bao%2C+Nan+Yang%2C+and+Ming+Zhou%2E+2020%2E+Minilm%3A+Deep+self%2Dattention+distillation+for+task%2Dagnostic+compression+of+pre%2Dtrained+transformers%2E+In+Proceedings+of+NeurIPS)
22. Xindi Wang, Mahsa Salmani, Parsa Omidi, Xiangyu Ren, Mehdi Rezagholizadeh, and Armaghan Eshaghi. 2024. [Beyond the limits: A survey of techniques to extend the context length in large language models](https://md.research.elamur.ai/free/en/papers/wang-2024-limits-survey-techniques-extend-context/). In Proceedings of IJCAI, pages 8299–8307.
23. Di Wu, Hongwei Wang, Wenhao Yu, Yuwei Zhang, Kai-Wei Chang, and Dong Yu. 2024. [Longmemeval: Benchmarking chat assistants on long-term interactive memory](https://md.research.elamur.ai/free/en/papers/wu-2024-longmemeval-benchmarking-chat-assistants-long/). ArXiv preprint, abs/2410.10813.
24. Wujiang Xu, Zujie Liang, Kai Mei, Hang Gao, Juntao Tan, and Yongfeng Zhang. 2025. [A-mem: Agentic memory for llm agents](https://md.research.elamur.ai/free/en/papers/xu-2025-mem-agentic-memory-llm-agents/). ArXiv preprint, abs/2502.12110.
25. An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, and 1 others. 2025. [Qwen3 technical report](https://md.research.elamur.ai/free/en/papers/yang-2025-qwen3-technical-report/). ArXiv preprint, abs/2505.09388.

*Updated 2 November 2026*