Researchers have proposed Spatial Memory Intelligence (SMI), described as the first framework to systematically employ an understanding model for spatial-memory management in long-video world models. The work targets a growing bottleneck: as memory sequences in long-video generation and world models become longer and more structurally complex, managing long-range spatial context becomes increasingly difficult.
World models predict future observations conditioned on user actions and historical memory, and have shown promise for interactive entertainment and embodied simulation. SMI builds on the advancing spatial reasoning capabilities of multimodal large language models (MLLMs) and the broader vision of unified models.
The framework introduces four coordinated atomic operations: spatial clustering, within-cluster sparsification, action-aware retrieval, and reliability-aware filtering. Together, these operations organize, sparsify, filter, and retrieve spatial memory.
According to the authors, SMI differs from prior work that addresses individual objectives such as memory compression, consistency, or stability. Instead, it tackles these challenges from a unified perspective, substantially reducing redundant memories while improving generation stability and spatial consistency.
The paper reports extensive experiments across multiple baselines, benchmarks, and world-model backbones, demonstrating effectiveness and generalizability with comprehensive improvements in memory sparsity, generation stability, and spatial consistency.
The authors motivate the approach by noting that spatial-memory management involves spatial perception and reasoning, tracking semantic changes, and understanding spatial relationships across time and viewpoints. They cite the strong spatial perception and reasoning capabilities demonstrated by GPT-6-Astra as reinforcing interest in whether spatial-memory management can improve as data and model size scale, and express hope that SMI is a step toward spatial-memory management within future unified models.