ARC: Addressable Recall Compaction for LLM Context Management
The Addressable Recall Compaction (ARC) framework distinguishes between archival storage and the presentation of active context in LLM agents. It maintains tool observations in a log that is append-only and ID-addressable, substituting older observations with concise citations. The agent can retrieve stored information using identifiers, eliminating the need to re-execute tools or depend exclusively on similarity-based retrieval. When tested on Qwen3-8B (16k context) and Qwen3-32B (32k context), ARC demonstrated impressive performance on the Needle-in-a-Haystack evaluation.
Key facts
- ARC is a context-management framework for long-horizon LLM agents.
- It separates archival storage from active-context presentation.
- Tool observations are stored in an append-only, ID-addressable log.
- Older observations are replaced with compact citations when compaction is required.
- Agents can request stored content using identifiers without re-executing tools.
- Evaluated on Qwen3-8B with 16k context window and Qwen3-32B with 32k context window.
- Evaluation uses the Needle-in-a-Haystack benchmark.
- ARC addresses limitations of discarding, summarizing, or retrieving earlier information.
Entities
—