Addressable Recall Compaction for Long Context-Window Control in AI Agents

2026-07-27

AI Summary

This work addresses the challenge faced by long-context large language model agents in effectively managing accumulated reasoning traces and tool observation data due to limited context windows, where existing compression methods often discard critical information. To overcome this, the authors propose the ARC framework, which decouples archived and active contexts using identifier-addressable logs, replaces raw content with compact references, and introduces an explicit, identifier-based on-demand recall mechanism—eliminating reliance on similarity-based retrieval or costly tool re-execution. Experiments on the Qwen3 model family demonstrate that ARC achieves 99.40% accuracy on Needle-in-a-Haystack tasks (versus 88.12% for the baseline) and improves performance on the LongBench-v2 Hard subset to 29.97% (from 28.25%), while significantly reducing service latency and HBM traffic.

Abstract

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.

Problem

• long context window
• context compaction
• AI agents
• information retention
• tool observations

Innovation

• Addressable Recall
• Context Compaction
• Long-context LLM Agents
• ID-addressable Log
• Efficient Retrieval