Long‑horizon language‑model agents continuously accumulate interaction histories, causing the context size to explode and making context compression essential for efficient execution. Most existing learned compression methods rely on feedback from downstream tasks or solver behavior, tying training to specific tasks or agents. SumOne trains a summarizer independently by using masked reconstruction. Given a summary of an expired history, a frozen reconstructor randomly masks portions of the original trajectory and tries to recover the masked content. The reconstruction error combined with a length penalty forms a task‑agnostic reward, which is optimized with GRPO. This reward is derived solely from raw trajectories, requiring no reference summaries or downstream task execution. At inference time only the lightweight summarizer is kept, enabling plug‑and‑play context compression for any unchanged agent. Systematic evaluation on long‑horizon search and software‑engineering tasks shows that SumOne reduces estimated inference cost by about 78 % on DeepSearchQA and 75 % on SWE‑bench, with an absolute error drop of 0.04 compared to using full context. The same checkpoint also transfers across downstream models and agent harnesses without additional fine‑tuning. Review: SumOne achieves task‑independent context compression through self‑supervised masked reconstruction, balancing compression ratio and information preservation, and demonstrates strong plug‑and‑play capability across diverse tasks.