Fetching from the wire…
01
02
03
04
05
06
07
08
Research2026-08-27 · source-backed
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.
Each link below shares sources, entities, or timing with this story.
Shared entity: Under / Same source domain / Shared topic / Earlier coverage / Tension
Both cover Under; reported by the same outlet (arxiv.org); overlapping topics (against, agent, baseline, choice, under).
Both cover Under; reported by the same outlet (arxiv.org); overlapping topics (budget, memory, under).
Both cover Under; reported by the same outlet (arxiv.org); overlapping topics (against, agent).
Shared entity: Under / Same source domain / Shared topic / Earlier coverage
Both cover Under; reported by the same outlet (arxiv.org); overlapping topics (budget, token, under).
Both cover Under; reported by the same outlet (arxiv.org); overlapping topics (agent, context, under).
Same source domain / Shared topic / Tension
Reported by the same outlet (arxiv.org); overlapping topics (ablation, agent, baseline, decision, memory); pushes against this story (versus).
Reported by the same outlet (arxiv.org); overlapping topics (agent, baseline, context, memory, token); pushes against this story (versus).
Shared entity: GLM / Same source domain / Shared topic / Tension
Both cover GLM; reported by the same outlet (arxiv.org); overlapping topics (agent, memory).