KOPE cut kernel-optimization token spend from 15.9B to 1.113B while raising pass rate from 60.0% to 84.6%.

The method stores hardware kernel optimization trajectories, with correctness and performance feedback, in an Experience Graph Memory that preserves decision order, observed outcomes, and abandoned branches, then retrieves under a fixed token budget (arXiv 2608.25570). Under the same GLM-5.2 setting, geometric-mean per-operator speedup is 1.54x the strongest baseline. The 53-operator ablation is the striking part: active context management alone drives both the pass-rate lift and the 14x token reduction against letting the agent build its own context. Preserving abandoned branches is the design choice most memory systems skip.