Event

ReSight: A Self-Improving Memory System for Embodied Agents

1 reports · 1 sources

Lirong Che and other authors released a paper proposing ReSight, a task-driven self-improving memory system for embodied agents. The system retains keyframes and their spatiotemporal associations, and separates observational evidence from extensible derived fields so that historical observations can be reprocessed according to new requirements. ReSight uses Explore-Distill-Mount (EDM) to identify capability gaps from execution and exploration records, and converts solutions obtained through exploration into reusable memory builders, query tools, and skills. The paper states that, within their respective benchmarks, with the execution model and online budget fixed and tools and skills frozen, one approach invokes registered tools within a limited online budget, while the other uses a programmable workspace and invests a larger exploration budget to inspect evidence, write programs, and acquire models as needed.