New AI memory system doubles materials science success rates

An arXiv paper argues that a lifelong AI partner for materials science can be built around persistent memory rather than any single agent. The proposed self-evolving framework stores scientific experience as inspectable facts and executable skills, allowing observations, failure boundaries, protocols, and validation checks to be retrieved, revised, and migrated across models. In 49 real-world materials-tool-use questions, this memory nearly doubled GPT-5.2 task success without updating model parameters. In equation-of-state calculations, the framework turned a wavefunction-initialization failure into a pre-execution guardrail, improving Correct/Partial/Error results from 22/1/4 to 25/2/0 and avoiding 92% of repeated errors. In 13 practical material simulation workflows, remembered skills and failure facts cut the trace burden in half and reduced tool calls by more than half by the third round. The authors suggest agent memory can act as a durable scientific asset—a portable, self-improving record that outlives any single model or agent stack.