AI agents that run for hours forget things. A new paper proposes fixing that with a filing system instead of a bigger brain.
Researchers describe ARC, short for Addressable Recall Compaction, a framework for managing the context window of long-running AI agents. Instead of summarizing or deleting old tool outputs when an agent runs low on context space, ARC stores every tool observation in an append-only log, each entry tagged with its own ID. When the active context fills up, older entries get swapped for short citations, and the agent can pull the original content back using that ID rather than re-running the tool or hoping a similarity search finds it. The team tested the approach on Qwen3-8B with a 16k-token window and Qwen3-32B with a 32k-token window.
On a needle-in-a-haystack recall test, ARC hit 99.40% exact-answer accuracy versus 88.12% for the best baseline in the comparison, while also cutting estimated serving time and memory traffic under the researchers' hardware-cost model. On the harder LongBench-v2 Hard subset, the gap narrowed to 29.97% versus 28.25%, a reminder that recall and reasoning are different problems.
Most context-compaction fixes bet on smarter summarization. ARC bets on better bookkeeping instead, a lower-risk wager that only solves half the long-context problem.