PEEK: Context Map Cache for Long-Context LLM Agents

PEEK, a novel system, offers a context map cache designed for large language model (LLM) agents that engage with extensive and recurring external contexts, including document collections and code repositories. Current approaches either track the agent's path, allow passive access to raw data, or focus on task-specific strategies, but none effectively retain reusable knowledge regarding the recurring context—such as its contents, organization, and historically valuable entities, constants, and schemas. By caching this orientation knowledge as a small, constant-sized element within the agent's prompt, PEEK ensures a continuous insight into the external context. The map is governed by a programmable cache policy featuring three components: a Distiller for extracting transferable knowledge, a Cartographer for map updates, and a third module (details not fully disclosed). This system is detailed in arXiv paper 2605.19932, submitted in 2025.