EvoScientist: Self-Evolving Multi-Agent AI Scientist with Persistent Memory Outperforms 7 State-of-the-Art Systems

arXiv·medium signalXBlueskyLinkedInCopy link

EvoScientist (arXiv:2603.08127) introduces a three-agent architecture — Researcher, Engineer, and Evolution Manager — where the EMA continuously distills successful strategies into two persistent memory modules: ideation memory (feasible research directions, with failed ones blacklisted) and experimentation memory (effective data processing and training strategies from past code search trajectories). Evaluated against 7 open-source and commercial AI scientist systems, EvoScientist achieves higher novelty, feasibility, relevance, and clarity scores on scientific idea generation benchmarks. The persistent-memory-over-static-pipeline pattern directly extends to production agent frameworks.