Search papers, labs, and topics across Lattice.
This paper introduces ARC (Addressable Recall Compaction), a novel context-management framework designed to enhance the performance of long-horizon LLM agents by effectively managing their accumulated reasoning traces and tool observations. By utilizing an append-only, ID-addressable log for archival storage, ARC allows agents to replace older observations with compact citations, enabling efficient retrieval without re-executing tools. The evaluation shows that ARC significantly outperforms existing compaction methods, achieving a 99.40% exact-answer accuracy on the Needle-in-a-Haystack task, while also reducing serving time and hardware memory traffic.
ARC achieves a staggering 99.40% accuracy in long-context tasks by revolutionizing how AI agents manage and recall information.
Long-horizon LLM agents accumulate reasoning traces, actions, and tool observations that can eventually exceed a model's fixed context window. Existing compaction methods address this limitation by discarding, summarizing, or retrieving earlier information, but they may remove task-critical details or fail to recover them reliably. We propose ARC (Addressable Recall Compaction), a context-management framework that separates archival storage from active-context presentation. ARC stores tool observations in an append-only, ID-addressable log and replaces older observations with compact citations when compaction is required. The agent can subsequently use these identifiers to request stored content without re-executing the corresponding tools or depending solely on similarity-based retrieval. We evaluate ARC using Qwen3-8B with a 16k context window and Qwen3-32B with a 32k context window. On the Needle-in-a-Haystack evaluation, ARC achieves an average exact-answer accuracy of 99.40%, compared with 88.12% for the best-performing baseline in our evaluation. ARC also reduces estimated serving time and HBM traffic under our hardware-cost model. On the LongBench-v2 Hard subset, ARC obtains an average accuracy of 29.97%, compared with 28.25% for the best-performing baseline. These results indicate that explicit, address-based recall can improve information retention and serving efficiency relative to the evaluated context-management baselines under the tested settings.
