What's this about?
Enhancing real-time streaming interactive world models with patch memory. 12 upvotes.
What is patch memory?
Remembering parts of the scene rather than whole frames. An interactive world model generating in real time cannot hold every previous frame, and summarising them loses spatial specificity. Patch memory retains regions — this corner of the room, this object's appearance — so that when the view returns there, the model recalls that specific patch rather than regenerating from a description. It is the same insight as LayerRecall from yesterday, at finer granularity.
Why is it worth noticing?
- Third long-term-memory-for-video paper in two days, after LayerRecall and Ring Forcing.
- "Real-time streaming interactive" is the hard combination — each constraint alone is manageable.
- Matrix-Game is on its 3.5 release, which means iteration rather than a first attempt.