# SUMONE: A summarizer for plug-and-play agent memory [Canonical reading](https://escapemeridian.com/readings/sumone-a-summarizer-for-plug-and-play-agent-memory-c11a1e77) Authors Jizhe Zhang, Ang Li, Chenheng Zhang and Yisen Wang introduce SumOne, a summarizer designed to compress language-model agents’ interaction histories without training it for a particular downstream task or agent. Their method uses masked reconstruction: a frozen reconstructor tries to recover randomly masked parts of the original history from its summary. Reconstruction quality, combined with a length penalty, rewards the summarizer through GRPO, using raw trajectories rather than reference summaries or downstream task execution. At inference time, the authors say, only the lightweight summarizer is needed, so agents can use compressed context without other changes. In evaluations on long-horizon search and software-engineering tasks, the authors report estimated inference-cost reductions of approximately 78% on DeepSearchQA and 75% on SWE-bench Verified, relative to full context, alongside absolute decreases of 0.04. They also report that the same checkpoint transfers across downstream models and agent harnesses without additional fine-tuning. As an Escape Meridian interpretation, the approach aims to reduce the cost of carrying long histories while avoiding task-specific training; the abstract does not establish whether the reported savings and transfer results hold beyond the evaluated settings. The abstract does not identify the metric for the absolute decreases. Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims. ## Original source arXiv cs.AI · Article Original source date: Nov 4, 2026 (2026-11-04). Date basis: Date basis not provided. Added to Escape Meridian: 2026-11-09T08:19:00.379Z. First observed: 2026-11-05T05:00:36.186Z. Retrieved: 2026-11-09T08:18:12.468Z. - [SUMONE: A Simple and Efficient Summarizer for Plug-and-Play Agent Memory](https://arxiv.org/abs/2611.06445) ## Limits The abstract does not identify the metric for the reported absolute decreases of 0.04. The abstract describes evaluations in long-horizon search and software-engineering tasks, but does not provide details about the evaluation setup or the range of downstream models and agent harnesses tested. ## Open question How well does SumOne’s reported transfer across downstream models and agent harnesses hold beyond the long-horizon search and software-engineering evaluations described in the abstract? ## Source-stated The authors introduce SumOne, a standalone summarizer trained with masked reconstruction from raw agent trajectories, without reference summaries or downstream task execution. > This yields self-supervision directly from raw trajectories, without reference summaries or downstream task execution. - [Supporting source](https://arxiv.org/abs/2611.06445) ## Source-stated The authors report estimated inference-cost reductions relative to full context on DeepSearchQA and SWE-bench Verified, alongside absolute decreases of 0.04. > SumOne reduces estimated inference cost by approximately 78\\% on DeepSearchQA and 75\\% on SWE-bench Verified relative to operating with full context, with absolute decreases of 0.04. - [Supporting source](https://arxiv.org/abs/2611.06445) ## Source-stated The authors report that the same checkpoint transfers across downstream models and agent harnesses without additional fine-tuning. > The same checkpoint also transfers across downstream models and agent harnesses without additional fine-tuning. - [Supporting source](https://arxiv.org/abs/2611.06445) ## Escape Meridian interpretation The authors’ method aims to lower the cost of carrying long interaction histories while avoiding task-specific training; its broader performance is not established by the abstract. > making context compaction essential for efficient execution. Existing learned compaction methods often rely on downstream task outcomes or solver behavior, coupling training to specific tasks or agents. - [Supporting source](https://arxiv.org/abs/2611.06445) ## Source and contributor context Zhang, Jizhe, Li, Ang, Zhang, Chenheng, Wang, Yisen · Article author These are credits stated by the source; they do not establish a profile match or authorship of every passage. ## Source media No suitable media was found by the supported page extraction. ## Terms and references Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement. ## Fine-tuning Further training of an existing model to adapt it to a particular task or use. - [Definition reference](https://developers.google.com/machine-learning/glossary#fine-tuning) ## arXiv arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed. - [Explore this context](/entities/arxiv) - [Reference](https://arxiv.org/) - [Reference](https://info.arxiv.org/help/api/index.html) ## TIME TIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services. - [Explore this context](/entities/time) - [Reference](https://time.com/) ## Editorial provenance Version 1. Publication ID: 25581d1c-85da-4df2-8bab-5575ba8aabb5. Prepared by gpt-6-luna; grounding review by gpt-6-luna. Source check: passed by anthropic/claude-opus-5.5 on 2026-11-09T08:19:00.379Z. This checks fidelity to the source, not whether its claims are independently true. Editorial selection: accept; reasons: material\_development. Observation: 3dd875acd865928803ecdadf4b218fda05956c17d7d164ddca36dc25e0822635. Source text hash: 6309fe84385b33a01c7bb576c719b2822e94756382fbd756034ff53c1a1d37e1. ## Revision and source-check history ### Version 1 — 2026-11-09T08:19:00.379Z Authors Jizhe Zhang, Ang Li, Chenheng Zhang and Yisen Wang introduce SumOne, a summarizer designed to compress language-model agents’ interaction histories without training it for a particular downstream task or agent. Their method uses masked reconstruction: a frozen reconstructor tries to recover randomly masked parts of the original history from its summary. Reconstruction quality, combined with a length penalty, rewards the summarizer through GRPO, using raw trajectories rather than reference summaries or downstream task execution. At inference time, the authors say, only the lightweight summarizer is needed, so agents can use compressed context without other changes. In evaluations on long-horizon search and software-engineering tasks, the authors report estimated inference-cost reductions of approximately 78% on DeepSearchQA and 75% on SWE-bench Verified, relative to full context, alongside absolute decreases of 0.04. They also report that the same checkpoint transfers across downstream models and agent harnesses without additional fine-tuning. As an Escape Meridian interpretation, the approach aims to reduce the cost of carrying long histories while avoiding task-specific training; the abstract does not establish whether the reported savings and transfer results hold beyond the evaluated settings. The abstract does not identify the metric for the absolute decreases. ## Original source arXiv cs.AI · Article Original source date: Nov 4, 2026 (2026-11-04). Date basis: Date basis not provided. Added to Escape Meridian: 2026-11-09T08:19:00.379Z. First observed: 2026-11-05T05:00:36.186Z. Retrieved: 2026-11-09T08:18:12.468Z. - [SUMONE: A Simple and Efficient Summarizer for Plug-and-Play Agent Memory](https://arxiv.org/abs/2611.06445) ## Limits The abstract does not identify the metric for the reported absolute decreases of 0.04. The abstract describes evaluations in long-horizon search and software-engineering tasks, but does not provide details about the evaluation setup or the range of downstream models and agent harnesses tested. ## Open question How well does SumOne’s reported transfer across downstream models and agent harnesses hold beyond the long-horizon search and software-engineering evaluations described in the abstract? ## Source-stated The authors introduce SumOne, a standalone summarizer trained with masked reconstruction from raw agent trajectories, without reference summaries or downstream task execution. > This yields self-supervision directly from raw trajectories, without reference summaries or downstream task execution. - [Supporting source](https://arxiv.org/abs/2611.06445) ## Source-stated The authors report estimated inference-cost reductions relative to full context on DeepSearchQA and SWE-bench Verified, alongside absolute decreases of 0.04. > SumOne reduces estimated inference cost by approximately 78\\% on DeepSearchQA and 75\\% on SWE-bench Verified relative to operating with full context, with absolute decreases of 0.04. - [Supporting source](https://arxiv.org/abs/2611.06445) ## Source-stated The authors report that the same checkpoint transfers across downstream models and agent harnesses without additional fine-tuning. > The same checkpoint also transfers across downstream models and agent harnesses without additional fine-tuning. - [Supporting source](https://arxiv.org/abs/2611.06445) ## Escape Meridian interpretation The authors’ method aims to lower the cost of carrying long interaction histories while avoiding task-specific training; its broader performance is not established by the abstract. > making context compaction essential for efficient execution. Existing learned compaction methods often rely on downstream task outcomes or solver behavior, coupling training to specific tasks or agents. - [Supporting source](https://arxiv.org/abs/2611.06445) ## Source and contributor context Zhang, Jizhe, Li, Ang, Zhang, Chenheng, Wang, Yisen · Article author These are credits stated by the source; they do not establish a profile match or authorship of every passage. ## Source media No suitable media was found by the supported page extraction. ## Terms and references Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement. ## Fine-tuning Further training of an existing model to adapt it to a particular task or use. - [Definition reference](https://developers.google.com/machine-learning/glossary#fine-tuning) ## arXiv arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed. - [Explore this context](/entities/arxiv) - [Reference](https://arxiv.org/) - [Reference](https://info.arxiv.org/help/api/index.html) ## TIME TIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services. - [Explore this context](/entities/time) - [Reference](https://time.com/) ## Editorial provenance Version 1. Publication ID: 25581d1c-85da-4df2-8bab-5575ba8aabb5. Prepared by gpt-6-luna; grounding review by gpt-6-luna. Source check: passed by anthropic/claude-opus-5.5 on 2026-11-09T08:19:00.379Z. This checks fidelity to the source, not whether its claims are independently true. Editorial selection: accept; reasons: material\_development. Observation: 3dd875acd865928803ecdadf4b218fda05956c17d7d164ddca36dc25e0822635. Source text hash: 6309fe84385b33a01c7bb576c719b2822e94756382fbd756034ff53c1a1d37e1. ### passed — 2026-11-09T08:19:00.379Z — anthropic/claude-opus-5.5 Publication: 25581d1c-85da-4df2-8bab-5575ba8aabb5 attribution — pass: The draft correctly names the four listed authors and attributes the method, results and transfer claims to them. Results are framed as reported findings, not established facts. > Authors:Jizhe Zhang, Ang Li, Chenheng Zhang, Yisen Wang summary — pass: The draft faithfully restates the abstract: a standalone summarizer, masked-reconstruction training with GRPO and a length penalty, and plug-and-play use. It reports the cost cuts and the unexplained absolute decreases of 0.04, flagging the missing metric. Plain-language framing (compressing agent histories to lower cost) keeps the central idea readable. Terms such as GRPO, checkpoint and agent harnesses are kept exactly and suit term-library tooltips. The question is a genuine open issue about generalization. > At inference time, only the lightweight summarizer is retained, enabling plug-and-play context compression for otherwise unchanged agents. date — pass: originalPublished 2026-11-04 matches the arXiv v1 submission date, not the later feed date. > \[Submitted on 4 Nov 2026\] support — pass: All source-stated claims match their verbatim excerpts. The labeled interpretation about lowering cost while avoiding task-specific training follows from the abstract's motivation. It does not claim real-world reliability. The limitations accurately note the unidentified metric and the sparse evaluation details. > Existing learned compaction methods often rely on downstream task outcomes or solver behavior, coupling training to specific tasks or agents. media — pass: The selected media is the verified direct arXiv abstract page for this paper, the most direct access to the source. ## Recorded references Explicit citations in public readings and public file attachments to this reading. Shared topics are not citations. Other artifact types need an explicit reference adapter. No public references recorded here yet. [Reference data (JSON)](/readings/c11a1e7777d16e40185284008e4ec1d75c985e35d0fe89c99b3c1ef85a407635/references.json) ## Corrections and challenges [Send a correction or an opposing source](https://escapemeridian.com/corrections/) [JSON record](https://escapemeridian.com/readings/sumone-a-summarizer-for-plug-and-play-agent-memory-c11a1e77.json)