PEEK: Context Map as an Orientation Cache for Long-Context LLM Agents

Researchers introduced PEEK, a persistent context-caching system for long-context LLM agents.

PEEK stores compact orientation knowledge in a fixed-size prompt artifact, maintained by a distiller, cartographer, and evictor.

In tests on long-context reasoning, information aggregation, and context learning, it beat strong baselines like ACE.

The gains held across different models and agent setups, while using fewer iterations and less cost.