Spatial Memory Intelligence

As historical memory grows, managing spatial memory becomes increasingly complex, requiring coordinated organization, maintenance, sparsification, and retrieval. This calls for a more intelligent and comprehensive memory-management approach.

We propose Spatial Memory Intelligence (SMI), the first understanding-driven unified spatial-memory manager. Four coordinated atomic operations bring semantic and spatial reasoning into the memory-management pipeline, achieving comprehensive improvements in memory sparsity, spatial consistency, and generation stability across multiple baselines, benchmarks, and world-model backbones.