PINNED

Self-Evolving Memory

Memory infrastructure that lets AI agents and robots keep learning from intermittent streams of experience. Instead of a dataset fixed before training, memory is an ongoing process of accumulation, self-organization, and recall, so agents can autonomously discover and recompose skills and adapt to new environments. We study its retrieval infrastructure, retrieval-augmented robot learning, methods for distilling experience and models, and integration with foundation models. JST PRESTO, principal investigator (2025—).