Long-running agents accumulate context while model memory stays fixed. This leads to a tradeoff: either discard older information or compress it. New work by @oneill_c explores repeated KV-cache compression for persistent agents using Attention Matching. Our research shows