ReflectWorld-MM: An Entity-Oriented Multimodal Memory System for Open-Ended Video Streams

ReflectWorld-MM proposes an entity-oriented multimodal memory architecture for assistants that continuously observe open-ended audiovisual streams. Its design combines bounded short-term perception with hierarchical long-term memory: multi-scale episodic memory, evolving entity-centric semantic memory, and procedural memory. The system is described as a complete implementation that can ingest arbitrary streams and connect to off-the-shelf assistants. According to the abstract, it achieves the best accuracy on all six evaluated long-video and lifelong-memory benchmarks, outperforming strong memory agents and a frontier model. Exact scores, datasets, baselines, and implementation details are not provided in the supplied summary.