---
title: "Memory as Action — paper by Zhang et al. (2026)"
description: "LLM agents on long tasks such as deep research or multi-step tool use accumulate large interaction… 21 references, cited by 8 papers here."
authors: ["Yuxiang Zhang", "Jiangming Shu", "Ye Ma", "Xueyuan Lin", "Shangxi Wu", "Jitao Sang"]
year: 2026
venue: "Findings of the Association for Computational Linguistics: ACL 2026"
source: "https://aclanthology.org/2026.findings-acl.956/"
doi: "10.18653/v1/2026.findings-acl.956"
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/zhang-2026-memory-as-action-autonomous-context/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/zhang-2026-memory-as-action-autonomous-context/)

[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/2/) / Memory as Action: Autonomous Context Curation for Long-Horizon Agentic Tasks

# Memory as Action: Autonomous Context Curation for Long-Horizon Agentic Tasks

**Authors:** [Yuxiang Zhang](https://md.research.elamur.ai/free/en/authors/yuxiang-zhang/), Jiangming Shu, Ye Ma, [Xueyuan Lin](https://md.research.elamur.ai/free/en/authors/xueyuan-lin/), Shangxi Wu and [Jitao Sang](https://md.research.elamur.ai/free/en/authors/jitao-sang/) · **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.956/) · [PDF](https://aclanthology.org/2026.findings-acl.956.pdf) · [DOI 10.18653/v1/2026.findings-acl.956](https://doi.org/10.18653/v1/2026.findings-acl.956) · [OpenAlex](https://openalex.org/W4415270781)

**Citations:** 2 citations (OpenAlex, 2026-10-14) · [8](https://md.research.elamur.ai/free/en/papers/zhang-2026-memory-as-action-autonomous-context/cited-by/) in our library, of the 149,439 papers whose bibliographies we read

## Summary
LLM agents on long tasks such as deep research or multi-step tool use accumulate large interaction histories. Irrelevant material dilutes attention and hurts reasoning. Most current fixes rely on heuristic triggers or external memory controllers that do not see the agent's reasoning state. MemAct moves context management into the agent's own policy. Alongside ordinary task actions like search, the model can call a Prune&Write operator. This operator removes earlier interaction turns by their IDs and adds a self-written note that summarizes, reflects, or plans. Both task performance and context use are then optimized together with end-to-end reinforcement learning.

Pruning breaks the usual assumption that context only grows, which makes standard RL training on such trajectories inconsistent. The authors address this with Dynamic Context Policy Optimization (DCPO), which splits trajectories into independent segments at each edit point. The trained 14B model reaches the accuracy of a model 16× larger while using about half the context. This suggests memory curation can be learned as a skill rather than hand-engineered.

## Key points
- Working memory is modeled as an MDP over turns, each with a unique ID. The memory action takes two arguments: a set of IDs to delete and new memory content, which is appended and stays editable later.
- DCPO segments trajectories at memory-edit points so training runs on standard RL infrastructure. The authors argue that simple attention masking cannot work, because deleted tokens have already shaped the key-value states of later tokens.
- MemAct-RL-14B matches Qwen3-235B accuracy while cutting average context length by 51%, and it also lowers end-to-end inference latency.
- Learned strategies depend on model size (7B vs 14B, SFT vs RL). The 14B-RL policy prunes in two typical sizes: about 2 records to drop noise during reasoning, and about 6 records to clear steps after finishing a sub-goal.
- Stated limitations include sparse final-outcome rewards, lossy summarization that cannot be undone, and uniform sampling of memory operations. Code and data are available at github.com/ADaM-BJTU/MemAct.

*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

- [Huawei Technologies](https://md.research.elamur.ai/free/en/companies/huawei-technologies/) — Company, Shenzhen, China · 96 works · 15,432 citations
- [Peng Cheng Laboratory](https://md.research.elamur.ai/free/en/labs/peng-cheng-laboratory/) — Research institute, Shenzhen, China · 47 works · 2,699 citations
- [The Hong Kong University of Science and Technology (Guangzhou)](https://md.research.elamur.ai/free/en/universities/the-hong-kong-university-of-science-and-technology-guangzhou/) — University, Guangzhou, China · 67 works · 1,765 citations
- [Beijing Jiaotong University](https://md.research.elamur.ai/free/en/universities/beijing-jiaotong-university/) — University, Beijing, China · 28 works · 1,754 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: Yuxiang Zhang, Jiangming Shu, Ye Ma, Xueyuan Lin, Shangxi Wu and Jitao Sang. 2026. Memory as Action: Autonomous Context Curation for Long-Horizon Agentic Tasks. 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{zhang2026memory,
  title = {{Memory as Action: Autonomous Context Curation for Long-Horizon Agentic Tasks}},
  author = {Yuxiang Zhang, Jiangming Shu, Ye Ma, Xueyuan Lin, Shangxi Wu and Jitao Sang},
  booktitle = {Findings of the Association for Computational Linguistics: ACL 2026},
  year = {2026},
  doi = {10.18653/v1/2026.findings-acl.956},
  url = {https://doi.org/10.18653/v1/2026.findings-acl.956}
}
```

- [← previous InfiGUIAgent: A Multimodal Generalist GUI Agent with Native Reasoning and Reflection](https://md.research.elamur.ai/free/en/papers/liu-2026-infiguiagent-multimodal-generalist-gui-agent/)
- [next → OneRec-Think: In-Text Reasoning for Generative Recommendation](https://md.research.elamur.ai/free/en/papers/liu-2026-onerec-think-text-reasoning-generative/)
- [↑ back to papers of 2026](https://md.research.elamur.ai/free/en/papers/2026/page/2/)

## References

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

1. Yucheng Li, Bo Dong, Frank Guerin, and Chenghua Lin. 2023. [Compressing context to enhance inference efficiency of large language models](https://md.research.elamur.ai/free/en/papers/li-2023-compressing-context-enhance-inference-efficiency/). EMNLP.
2. Zijian Chen, Xueguang Ma, Shengyao Zhuang, Ping Nie, Kai Zou, Andrew Liu, Joshua Green, Kshama Patel, Ruoxi Meng, Mingyi Su, Sahel Sharifymoghaddam, Yanxi Li, Haoran Hong, Xinyu Shi, Xuye Liu, Nandan Thakur, Crystina Zhang, Nelson F. Liu, Kevin Lin, John Hewitt, Ashwin Paranjape, Michele Bevilacqua,… — [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/)
3. Miao Lu, Weiwei Sun, Weihua Du, Zhan Ling, Xuesong Yao, Kang Liu, and Jiecao Chen. 2025. Scaling llm multi-turn rl with end-to-end summarization-based context management. arXiv preprint arXiv:2510.06727. — [arXiv:2510.06727](https://arxiv.org/abs/2510.06727) [find](https://scholar.google.com/scholar?q=Miao+Lu%2C+Weiwei+Sun%2C+Weihua+Du%2C+Zhan+Ling%2C+Xuesong+Yao%2C+Kang+Liu%2C+and+Jiecao+Chen%2E+2025%2E+Scaling+llm+multi%2Dturn+rl+with+end%2Dto%2Dend+summarization%2Dbased+context+management%2E+arXiv+preprint+arXiv%3A2510%2E067)
4. Lingrui Mei, Jiayu Yao, Yuyao Ge, Yiwei Wang, Baolong Bi, Yujun Cai, Jiazhi Liu, Mingyu Li, Zhong-Zhi Li, Duzhen Zhang, Chenlin Zhou, Jiayi Mao, Tianze Xia, Jiafeng Guo, and Shenghua Liu. 2025. A survey of context engineering for large language models. arXiv preprint arXiv:2507.13334. — [arXiv:2507.13334](https://arxiv.org/abs/2507.13334) [find](https://scholar.google.com/scholar?q=Lingrui+Mei%2C+Jiayu+Yao%2C+Yuyao+Ge%2C+Yiwei+Wang%2C+Baolong+Bi%2C+Yujun+Cai%2C+Jiazhi+Liu%2C+Mingyu+Li%2C+Zhong%2DZhi+Li%2C+Duzhen+Zhang%2C+Chenlin+Zhou%2C+Jiayi+Mao%2C+Tianze+Xia%2C+Jiafeng+Guo%2C+and+Shenghua+Liu%2E+2025%2E+A+surv)
5. OpenAI. 2025. [gpt-oss-120b & gpt-oss-20b model card](https://md.research.elamur.ai/free/en/papers/agarwal-2025-gpt-oss-120b-gpt-oss/). arXiv preprint arXiv:2508.10925.
6. Charles Packer, Sarah Wooders, Kevin Lin, Vivian Fang, Shishir G. Patil, Ion Stoica, and Joseph E. Gonzalez. 2023. Memgpt: Towards llms as operating systems. arXiv preprint arXiv:2310.08560. [find](https://scholar.google.com/scholar?q=Charles+Packer%2C+Sarah+Wooders%2C+Kevin+Lin%2C+Vivian+Fang%2C+Shishir+G%2E+Patil%2C+Ion+Stoica%2C+and+Joseph+E%2E+Gonzalez%2E+2023%2E+Memgpt%3A+Towards+llms+as+operating+systems%2E+arXiv+preprint+arXiv%3A2310%2E08560%2E)
7. Bowen Peng, Jeffrey Quesnelle, Honglu Fan, and Enrico Shippole. 2024. Yarn: Efficient context window extension of large language models. ICLR. [find](https://scholar.google.com/scholar?q=Bowen+Peng%2C+Jeffrey+Quesnelle%2C+Honglu+Fan%2C+and+Enrico+Shippole%2E+2024%2E+Yarn%3A+Efficient+context+window+extension+of+large+language+models%2E+ICLR%2E)
8. Ofir Press, Muru Zhang, Sewon Min, Ludwig Schmidt, Noah A Smith, and Mike Lewis. 2023. Measuring and narrowing the compositionality gap in language models. EMNLP. [find](https://scholar.google.com/scholar?q=Ofir+Press%2C+Muru+Zhang%2C+Sewon+Min%2C+Ludwig+Schmidt%2C+Noah+A+Smith%2C+and+Mike+Lewis%2E+2023%2E+Measuring+and+narrowing+the+compositionality+gap+in+language+models%2E+EMNLP%2E)
9. Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, and 1 others. 2024. [Deepseekmath: Pushing the limits of mathematical reasoning in open language models](https://md.research.elamur.ai/free/en/papers/shao-2024-deepseekmath-pushing-limits-mathematical-reasoning/). arXiv preprint arXiv:2402.03300.
10. Qwen Team and 1 others. 2024. Qwen2 technical report. arXiv preprint arXiv:2407.10671. [find](https://scholar.google.com/scholar?q=Qwen+Team+and+1+others%2E+2024%2E+Qwen2+technical+report%2E+arXiv+preprint+arXiv%3A2407%2E10671%2E)
11. Tongyi DeepResearch Team, Baixuan Li, Bo Zhang, Dingchu Zhang, Fei Huang, Guangyu Li, Guoxin Chen, Huifeng Yin, Jialong Wu, Jingren Zhou, and 1 others. 2025. Tongyi deepresearch technical report. arXiv preprint arXiv:2510.24701. — [arXiv:2510.24701](https://arxiv.org/abs/2510.24701) [find](https://scholar.google.com/scholar?q=Tongyi+DeepResearch+Team%2C+Baixuan+Li%2C+Bo+Zhang%2C+Dingchu+Zhang%2C+Fei+Huang%2C+Guangyu+Li%2C+Guoxin+Chen%2C+Huifeng+Yin%2C+Jialong+Wu%2C+Jingren+Zhou%2C+and+1+others%2E+2025%2E+Tongyi+deepresearch+technical+report%2E+arXi)
12. 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/). NeurIPS.
13. Sikuan Yan, Xiufeng Yang, Zuchao Huang, Ercong Nie, Zifeng Ding, Zonggen Li, Xiaowen Ma, Hinrich Schütze, Volker Tresp, and Yunpu Ma. 2025. [Memory-r1: Enhancing large language model agents to manage and utilize memories via reinforcement learning](https://md.research.elamur.ai/free/en/papers/yan-2026-memory-r1-enhancing-large-language/). arXiv preprint arXiv:2508.19828.
14. Zhilin Yang, Peng Qi, Saizheng Zhang, Yoshua Bengio, William W Cohen, Ruslan Salakhutdinov, and Christopher D Manning. 2018. [Hotpotqa: A dataset for diverse, explainable multi-hop question answering](https://md.research.elamur.ai/free/en/papers/yang-2018-hotpotqa-dataset-diverse-explainable-multi/). ACL.
15. Hongli Yu, Tinghong Chen, Jiangtao Feng, Jiangjie Chen, Weinan Dai, Qiying Yu, Ya-Qin Zhang, Wei-Ying Ma, Jingjing Liu, Mingxuan Wang, and 1 others. 2025. Memagent: Reshaping long-context llm with multi-conv rl-based memory agent. arXiv preprint arXiv:2507.02259. — [arXiv:2507.02259](https://arxiv.org/abs/2507.02259) [find](https://scholar.google.com/scholar?q=Hongli+Yu%2C+Tinghong+Chen%2C+Jiangtao+Feng%2C+Jiangjie+Chen%2C+Weinan+Dai%2C+Qiying+Yu%2C+Ya%2DQin+Zhang%2C+Wei%2DYing+Ma%2C+Jingjing+Liu%2C+Mingxuan+Wang%2C+and+1+others%2E+2025%2E+Memagent%3A+Reshaping+long%2Dcontext+llm+with+mul)
16. Zhenyu Zhang, Ying Sheng, Tianyi Zhou, Tianlong Chen, Lianmin Zheng, Ruisi Cai, Zhao Song, Yuandong Tian, Christopher Re, Clark Barrett, Zhangyang Wang, and Beidi Chen. 2023. H2o: Heavy-hitter oracle for efficient generative inference of large language models. NeurIPS. [find](https://scholar.google.com/scholar?q=Zhenyu+Zhang%2C+Ying+Sheng%2C+Tianyi+Zhou%2C+Tianlong+Chen%2C+Lianmin+Zheng%2C+Ruisi+Cai%2C+Zhao+Song%2C+Yuandong+Tian%2C+Christopher+Re%2C+Clark+Barrett%2C+Zhangyang+Wang%2C+and+Beidi+Chen%2E+2023%2E+H2o%3A+Heavy%2Dhitter+oracle+)
17. Lianmin Zheng, Liangsheng Yin, Zhiqiang Xie, Chuyue Sun, Jeff Huang, Cody Hao Yu, Shiyi Cao, Christos Kozyrakis, Ion Stoica, Joseph E. Gonzalez, Clark W. Barrett, and Ying Sheng. 2024. [Sglang: Efficient execution of structured language model programs](https://md.research.elamur.ai/free/en/papers/zheng-2023-sglang-efficient-execution-structured-language/). NeurIPS.
18. Zijian Zhou, Ao Qu, Zhaoxuan Wu, Sunghwan Kim, Alok Prakash, Daniela Rus, Jinhua Zhao, Bryan Kian Hsiang Low, and Paul Pu Liang. 2025. Mem1: Learning to synergize memory and reasoning for efficient long-horizon agents. arXiv preprint arXiv:2506.15841. — [arXiv:2506.15841](https://arxiv.org/abs/2506.15841) [find](https://scholar.google.com/scholar?q=Zijian+Zhou%2C+Ao+Qu%2C+Zhaoxuan+Wu%2C+Sunghwan+Kim%2C+Alok+Prakash%2C+Daniela+Rus%2C+Jinhua+Zhao%2C+Bryan+Kian+Hsiang+Low%2C+and+Paul+Pu+Liang%2E+2025%2E+Mem1%3A+Learning+to+synergize+memory+and+reasoning+for+efficient+lo)
19. Harsh Trivedi, Niranjan Balasubramanian, Tushar Khot, and Ashish Sabharwal. 2022. Musique: Multi-hop questions via single-hop question composition. Transactions of the Association for Computational Linguistics. [find](https://scholar.google.com/scholar?q=Harsh+Trivedi%2C+Niranjan+Balasubramanian%2C+Tushar+Khot%2C+and+Ashish+Sabharwal%2E+2022%2E+Musique%3A+Multi%2Dhop+questions+via+single%2Dhop+question+composition%2E+Transactions+of+the+Association+for+Computational+Li)
20. Jason Wei, Zhiqing Sun, Spencer Papay, Scott McKinney, Jeffrey Han, Isa Fulford, Hyung Won Chung, Alex Tachard Passos, William Fedus, and Amelia Glaese. 2025. [Browsecomp: A simple yet challenging benchmark for browsing agents](https://md.research.elamur.ai/free/en/papers/wei-2025-browsecomp-simple-yet-challenging-benchmark/). arXiv preprint arXiv:2504.12516.
21. Xixi Wu, Kuan Li, Yida Zhao, Liwen Zhang, Litu Ou, Huifeng Yin, Zhongwang Zhang, Xinmiao Yu, Dingchu Zhang, Yong Jiang, and 1 others. 2025. [Resum: Unlocking long-horizon search intelligence via context summarization](https://md.research.elamur.ai/free/en/papers/wu-2025-resum-unlocking-long-horizon-search/). arXiv preprint arXiv:2509.13313.

*Updated 2 November 2026*