SumOne: A Plug-and-Play Memory Summarizer for Agents

SumOne uses masked reconstruction plus GRPO to train a task-agnostic summarizer, drastically compressing agent context with almost no performance loss.
Long-horizon LLM agents accumulate extensive interaction history, causing context bloat and reduced efficiency; existing compression methods depend on downstream task results or solver behavior and are coupled to specific tasks and agents. Researchers propose SumOne, a standalone summarizer trained via masked reconstruction: the summarizer compresses stale history, a frozen reconstructor recovers randomly masked trajectory portions, and reconstruction quality plus a length penalty forms a task-agnostic reward optimized directly by GRPO without reference summaries or downstream execution. At inference only the lightweight summarizer is kept, making it plug-and-play. On DeepSearchQA and SWE-bench Verified it cuts inference cost by ~78% and ~75% versus full context, with only a 0.04 absolute performance drop, and the same checkpoint transfers across downstream models and agent frameworks.
Title
SumOne: A Plug-and-Play Memory Summarizer for Agents
Meta Description
SumOne is a plug-and-play memory summarizer for agents. Trained with masked reconstruction and GRPO, it cuts inference cost by ~78% on DeepSearchQA and ~75% on SWE-bench Verified, with only a 0.04 drop in performance, and transfers across models and frameworks.
Article Content
Long-horizon language model agents accumulate large amounts of interaction history during operation, causing context bloat that hurts execution efficiency. Existing context compression methods typically rely on downstream task results or solver behavior, coupling training to a specific task or agent. To address this, researchers propose SumOne, a standalone summarizer that preserves information in agent history through masked-reconstruction training. Given a summary of stale history, a frozen reconstructor recovers randomly masked portions of the original trajectory; reconstruction quality combined with a length penalty provides a task-agnostic reward for optimizing the summarizer with GRPO. This self-supervised approach learns directly from raw trajectories without reference summaries or downstream task execution. At inference time, only the lightweight summarizer is kept, enabling plug-and-play context compression without altering the original agent. Systematic evaluation on long-horizon search and software engineering tasks shows that SumOne reduces inference cost by about 78% on DeepSearchQA and about 75% on SWE-bench Verified (compared with full context), with an absolute performance drop of only 0.04. The same checkpoint transfers across downstream models and agent frameworks without additional fine-tuning.

Core Method: Masked Reconstruction and GRPO Optimization
SumOne's core innovation is decoupling summarizer training from downstream tasks. Traditional methods rely on task feedback to optimize summaries, whereas SumOne uses only the raw trajectory itself: for a stale history segment, the summarizer generates a summary, and then a frozen reconstructor attempts to recover randomly masked portions of the original trajectory from the summary. Higher reconstruction quality indicates that the summary retains more complete information; a length penalty is also added to prevent overly long summaries. This reward signal directly optimizes the summarizer via GRPO, without any reference summary or task execution.
This self-supervised approach enables SumOne to learn from massive amounts of unlabeled agent interaction data, and the trained summarizer generalizes well. In experiments, the same checkpoint is effective across different downstream models and agent frameworks without fine-tuning, demonstrating its plug-and-play nature.
GRPO (Group Relative Policy Optimization) is a reinforcement learning algorithm for large language model policy optimization, first proposed by the DeepSeekAd team as a replacement for the separately trained critic network in PPO. It samples a group of candidate outputs for the same input, normalizes each sample's reward using the mean and standard deviation of rewards within the group to obtain relative advantages, and then updates the policy accordingly. This within-group relative comparison eliminates the value network, significantly reducing training memory and compute overhead, and is insensitive to reward scale. In SumOne, the summarizer is treated as the policy model, and the reward composed of reconstruction quality and length penalty is normalized within the group to directly drive summarizer updates, so training requires neither a differentiable reconstruction objective nor human-annotated reference summaries.
Masked reconstruction draws on the idea of BERT-style masked language modeling: random tokens in the original trajectory are masked, and the model must recover these masked contents using only the summary. It constitutes an information-bottleneck-style self-supervised signal—if the summary omits key information, the reconstructor cannot restore the masked segments, and the reward decreases accordingly. Thus reconstruction fidelity can serve as a proxy for the information completeness of the summary, without the need to manually write reference summaries or run downstream tasks to evaluate summary quality.
Performance Evaluation: Substantially Lower Inference Cost
The researchers systematically evaluated SumOne on long-horizon search and software engineering tasks. On the DeepSearchQA dataset, SumOne reduces estimated inference cost by about 78% compared with using full context; on SWE-bench Verified, it reduces cost by about 75%. At the same time, the absolute performance drop is only 0.04, which is almost negligible. This means agents can significantly reduce computational overhead while maintaining nearly the same performance.
Interestingly, this cost reduction does not come at the expense of task success rate. Because the summarizer retains only a lightweight structure, inference only needs to load the summarizer, and the original agent requires no modification to achieve context compression. This is especially valuable for tasks that need to handle long histories (such as multi-turn search and code repair).
The DeepSearchQA and SWE-bench Verified mentioned in the article represent two types of long-horizon agent evaluation. DeepSearchQA targets deep-search question answering, requiring agents to perform multi-turn retrieval, cross-validation, and synthesis of multi-source information before answering; it involves many interaction turns and fast context growth, making it a typical scenario for testing long-history compression. SWE-bench Verified is a human-verified subset of SWE-bench, starting from issues in real GitHub repositories and requiring agents to locate bugs, modify code, and pass hidden tests; tasks span extensive file reading, command execution, and trial-and-error iteration, with extremely long trajectories. Together they cover both "information-retrieval-intensive" and "tool-call-intensive" long-horizon workloads, so reporting both cost and performance on these two benchmarks can comprehensively demonstrate the generality of the compression method. Note that the "about 78%/75% inference cost reduction" in the article refers to estimated cost compared with the full-context baseline, while "absolute performance drop of only 0.04" indicates that compression barely harms task success rate.
Plug-and-Play and Cross-Framework Transfer
Another advantage of SumOne is plug-and-play integration. Agents do not need to change their own architecture or training pipeline; they only need to connect the SumOne summarizer in the context management stage to obtain compressed history representations. In addition, the same checkpoint shows good transferability across different downstream models (such as different base language models) and different agent frameworks (such as different tool-calling approaches), without additional fine-tuning. This lowers the deployment barrier, allowing SumOne to be widely used as a general context compression component in various agent systems.
Conclusion and Outlook
Through masked reconstruction and GRPO optimization, SumOne achieves task-agnostic training of agent history summarization, significantly reducing inference cost while maintaining performance. Its plug-and-play nature makes it easy to integrate into existing agent systems. In the future, this self-supervised summarization method is expected to further extend to more types of agent tasks and longer time spans, providing a foundation for building efficient and sustainable agent systems.
One-line Summary
SumOne trains a task-agnostic summarizer with masked reconstruction and GRPO, drastically compressing agent context with almost no performance loss.
Paragraph Summary
Long-horizon language model agents accumulate large amounts of interaction history during operation, causing context bloat and reduced execution efficiency; existing compression methods mostly rely on downstream task results or solver behavior and are coupled to specific tasks and agents. Researchers propose SumOne—a standalone summarizer that preserves historical information through masked-reconstruction training: after the summarizer compresses stale history, a frozen reconstructor attempts to recover randomly masked portions of the original trajectory; reconstruction quality combined with a length penalty forms a task-agnostic reward that directly optimizes the summarizer via GRPO, without reference summaries or downstream execution. At inference time, only the lightweight summarizer is kept, making it plug-and-play. On DeepSearchQA and SWE-bench Verified, it reduces inference cost by about 78% and 75% respectively compared with full context, with an absolute performance drop of only 0.04, and the same checkpoint can transfer across downstream models and agent frameworks.
Related articles

OpenAI to Shut Down Agent Builder and Two Other Services on November 30, Developers Must Migrate Urgently
OpenAI will shut down Agent Builder, the evals dashboard, and reusable prompts on November 30; evals is already read-only. Here's the scope, OpenAI's migration paths (Agent SDK, PromptFu, your own code), and what developers should export and rebuild this week.

RankSpot: A Pay-Per-Use SEO & AI Search Execution Layer That Lets Agents Write, Publish, and Build Backlinks for You
RankSpot is an SEO & AI search execution layer for AI Agents. Connect via MCP to Claude or ChatGPT to write, publish & build backlinks. Pay-per-use, $3/article.

Appxiom: Giving AI Coding Assistants Production Context
Appxiom gives AI coding agents production context—crashes, jank, memory leaks, failed network calls—across Android, iOS, Flutter, KMP, and Web, ranking bugs by Goal Friction Impact.