ARC Improves LLM Context Window Management for AI Agents.

Large Language Model (LLM) agents operating over long horizons often accumulate vast amounts of information—reasoning traces, actions, and tool observations—that can quickly exceed the model's fixed context window. Traditional compaction methods, which involve discarding, summarizing, or retrieving information, frequently risk losing critical details or failing to recover them reliably. The proposed framework, Addressable Recall Compaction (ARC), offers a novel solution by decoupling archival storage from the active context. ARC stores tool observations in an append-only, ID-addressable log. When context compaction is necessary, older observations are replaced with compact citations. This allows the agent to explicitly request and retrieve any stored content using its unique identifier, bypassing the need for re-execution or sole reliance on similarity-based retrieval. Evaluations showed ARC achieving 99.40% exact-answer accuracy on the Needle-in-a-Haystack benchmark, outperforming baselines, and also demonstrating improved serving time and HBM traffic efficiency.