ReflectWorld-MM stores video memory around entities, not frames, and tops 6 benchmarks

Xiaokang Ma · hf · 2026-07-21

ReflectWorld-MM proposes an entity-oriented multimodal memory system for open-ended video streams.

It addresses the limits of existing video-memory systems, which usually store memories in the model context or a flat feature store and organize them around frames rather than persistent entities.
The system combines three components: a perception front end that converts audiovisual streams into entity-resolved observations, a hierarchical long-term memory inspired by human memory theory, and a full implementation that can ingest arbitrary streams and plug into off-the-shelf assistants.
Its long-term memory includes multi-scale episodic memory, evolving entity-centric semantic memory, and procedural memory.
Across six long-video and lifelong-memory benchmarks, the paper reports best accuracy on all six, beating strong memory agents and a frontier model.

More from Multimodal