LOCI Uses Spatial Linear Memory for Streaming World Models
September 30, 2026
LOCI implements a hybrid architecture for video world models, alternating between transformer blocks with key-value caches and blocks with recurrent linear-attention memory. The recurrent component uses projective camera geometry to improve retrieval of past observations.
HOW THIS AFFECTS YOU
●
builderYou can build streaming models that retain high-resolution detail via KV caches while maintaining compact long-term memory.
●
researcherThis integrates geometric priors into recurrent memory for spatial-temporal modeling.