ACON Context Compression: Failure-Driven Guideline Optimization Reduces Agent Memory 26-54% While Preserving 95%+ Task Accuracy — Works with Closed-Source Models

OpenReview·high signal

ACON (Agent Context Optimization) is a new framework from ICLR that compresses both environment observations and interaction histories for long-horizon LLM agents. The approach analyzes paired trajectories where full context succeeds but compressed context fails, then updates compression guidelines in natural language. Results on AppWorld, OfficeBench, and Multi-objective QA show 26-54% peak token reduction with 95%+ accuracy preserved. Critically, the approach is gradient-free and works with closed-source models like Claude and GPT. Distilled compressors can boost smaller LMs by up to 46%.