Spatial-FWP Mamba - a dual-memory architecture for spatial sequence tasks.

Abstract

Spatial-FWP Mamba is a dual-memory architecture for spatial sequence tasks. Mamba maintains temporal context, such as recent events and task progress, while Spatial-FWP maintains spatial context associated with previously visited locations. The two memory streams are updated independently. Spatial information retrieved from Spatial-FWP is not continuously written back into the recurrent state of Mamba. Instead, temporal and spatial context are combined only when the model performs the current prediction or decision. This helps reduce long-term interference between spatial memory and temporal state.