Dropping old reasoning blocks cut cache-read tokens 33.3% while reward went up.

arXiv 2609.29875 is training-free and online: rank past reasoning blocks by frozen-proxy entropy, delete them, keep every action, tool call and observation. On 260 WorkBuddyBench tasks average reward rose from 0.699 to 0.718, with input, output and cache-read tokens down 25.5%, 14.4% and 33.3%. Probing suggests old reasoning becomes safe to drop once its derived state has been written out to files, code or tool output. Which is an argument for designing agents that externalize state to disk, since doing so makes their own context compressible.