ReflectWorld-MM: An Entity-Oriented Multimodal Memory System for Open-Ended Video Streams
The system makes long video memory revolve around recurring entities, not frames.
ReflectWorld-MM turns audiovisual streams into entity-resolved observations, then stores them in a hierarchical memory split across episodic, semantic, and procedural components. The paper says this is meant for open-ended streams where people or objects may disappear and return over time. In six long-video and lifelong-memory benchmarks, it reports the best accuracy on all six, ahead of other memory agents and a frontier model. HF Daily Papers' note
ReflectWorld-MM turns audiovisual streams into entity-resolved observations, then stores them in a hierarchical memory split across episodic, semantic, and procedural components. The paper says this is meant for open-ended streams where people or objects may disappear and return over time. In six long-video and lifelong-memory benchmarks, it reports the best accuracy on all six, ahead of other memory agents and a frontier model. HF Daily Papers' note
score 4