Genie 3 can generate dynamic worlds that you can navigate in real time at 24 frames per second, retaining consistency for a few minutes, with visual memory extending as far back as one minute ago. The blog explicitly states that the model can currently support a few minutes of continuous interaction, rather than extended hours, which is listed as one of the current limitations of the system.
The “few minutes” interaction window represents a significant technical achievement in world model consistency. Maintaining coherent environments over time is computationally challenging because generating an environment auto-regressively is generally a harder technical problem than generating an entire video, since inaccuracies tend to accumulate over time. Despite this challenge, Genie 3 manages to keep environments physically consistent throughout the interaction period, ensuring that objects, lighting, and spatial relationships remain believable as users explore.
The one-minute visual memory capability means that if you explore a location and then move away, you can return to that same location and find it rendered consistently with how it appeared before. This is crucial for creating immersive experiences where users feel they’re exploring a persistent world rather than randomly generated scenes. While the total interaction time is currently limited to a few minutes, this duration is sufficient for meaningful exploration, experimentation with different scenarios, and testing of AI agents in simulated environments. The limitation appears to be primarily technical, related to computational requirements and memory constraints of maintaining long-term consistency in auto-regressive generation.