Yes, Genie 3 demonstrates both object permanence and memory of past interactions, though within the constraints of its few-minute interaction window. Visual memory extending as far back as one minute ago enables the system to maintain consistency when users revisit locations. The blog provides a specific example: The trees to the left of the building remain consistent throughout the interaction, even as they go in and out of view. This demonstrates that objects maintain their properties, positions, and appearance even when temporarily outside the user’s field of view.
The system achieves this through its auto-regressive generation approach, where the model has to take into account the previously generated trajectory that grows with time. For example, if the user is revisiting a location after a minute, the model has to refer back to the relevant information from a minute ago. This technical implementation means that Genie 3 isn’t just generating new content randomly—it’s actively maintaining a coherent representation of the world state and referencing previous frames to ensure consistency when rendering familiar areas.
However, it’s important to note that this memory capability is constrained by the overall interaction duration. Since the system currently supports only a few minutes of continuous interaction, the object permanence and memory features operate within this timeframe. The one-minute visual memory window represents the practical limit of how far back the system can reliably reference previous states while maintaining real-time performance. This limitation is likely due to computational constraints, as maintaining longer-term memory would require significantly more processing power and storage to track all the objects, environmental states, and spatial relationships across extended interaction sessions.