Cache Engineering for Agents

Researchers propose cache engineering as a method for LLM agents to directly control their active KV cache without modifying the trajectory or re-prefilling retained context, using drop to exclude selected cache from future attention and repos to adjust retained keys' position encodings.