Persistent agent memory nearly doubled task success and cut tokens 50% by round three — with no model updates

arXiv 2608.11224·medium signalXBlueskyLinkedInCopy link

This work argues the durable asset is the memory, not the agent implementation: a self-evolving framework stores experience as inspectable facts and executable skills that survive model swaps. On 49 real-world materials tool-use questions spanning 138 subtasks, memory nearly doubled task success with zero parameter updates; equation-of-state outcomes moved from 22 correct / 1 partial / 4 error to 25 / 2 / 0, avoiding 92% of repeated errors. Across 13 simulation workflows the aggregate token burden halved and tool calls dropped more than 2x by the third round — the clearest number yet on how fast a failure-fact store pays for itself.