Long-horizon language-model agents keep building up reasoning history as they work. Even after earlier decisions have been carried out and their outcomes observed, that history stays in context, steadily increasing context length and inference cost.
The paper, titled "When Can Agents Forget Their Reasoning? ICLR for Long-Horizon Agent Context Compression," asks when agents can safely forget that reasoning. The abstract explicitly contrasts this with static chain-of-thought compression, suggesting the authors see agent context compression as a distinct problem rather than a direct extension of prior work.
The source is an arXiv abstract only, so the proposed compression mechanism is not described in detail. The paper is listed as arXiv:2609.29875v1.