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

As AI agents perform increasingly long-horizon tasks, they accumulate reasoning traces and tool observations that eventually exceed the model's fixed context window. This leads to "context-window overflow," where the agent can no longer process new information. Current compaction methods—such as truncation, summarization, or semantic retrieval—are inherently lossy, meaning critical details are often permanently discarded or paraphrased, leading to agent failure. This paper asks: can we maintain a bounded active context while ensuring that all historical information remains exactly and reliably recoverable?

To solve this, the authors introduce Addressable Recall Compaction (ARC). ARC decouples the agent's "active view" (what the model sees) from its "archival storage" (what the system keeps).

1. Append-only Store: Every tool observation is stored verbatim in an ID-addressable log using a hash-based identifier (e.g., §a91f3c20).
2. Bounded Active View: When the transcript exceeds the context limit, ARC replaces older, verbose observations with compact citations.
3. Explicit Recall: If the agent needs the original information, it issues a _recall §id command. The framework intercepts this, fetches the exact content from the store, and injects it back into the context window.

This approach ensures the system is "observation-lossless" at the architectural level, as no data is ever deleted, only moved out of the immediate active view.

The authors evaluated ARC using Qwen3-8B and Qwen3-32B models on the "Needle-in-a-Haystack" benchmark and the LongBench-v2 hard subset. ARC significantly outperformed existing baselines in both accuracy and reliability. On the Needle-in-a-Haystack task, ARC achieved an average accuracy of 99.40%, compared to 88.12% for the best-performing baseline. Furthermore, by keeping the active context window bounded, ARC reduced estimated hardware serving costs and HBM traffic, demonstrating that explicit, address-based recall is more efficient and accurate than semantic-based retrieval or lossy summarization.