LongSpace: Exploring Long-Horizon Spatial Memory from Perception to Recall in Video

Researchers introduce LongSpace-Bench, a video benchmark for evaluating multimodal AI models' ability to remember and retrieve spatial information across long videos, and propose LongSpace, a memory framework that improves long-horizon spatial reasoning by incorporating 3D structural cues and layer-aware memory retrieval.

LongSpace addresses a critical limitation in current multimodal large language models: their ability to maintain spatial coherence and memory across extended video sequences. While MLLMs have improved significantly in processing longer visual inputs, they struggle with tasks requiring persistent spatial recall—a capability essential for autonomous systems. The research tackles real-world challenges in autonomous driving and robotic navigation, where models must track previously observed layouts, routes, and object states rather than simply analyzing current frames.

The introduction of LongSpace-Bench provides the AI research community with a standardized evaluation framework focused on room-tour videos, enabling more rigorous assessment of spatial memory capabilities. This benchmark covers three critical dimensions: scene perception, spatial relations, and spatial memory retrieval. The proposed LongSpace framework itself demonstrates technical innovation by segmenting long videos into sequential chunks, integrating 3D structural information early in processing, and implementing question-guided memory retrieval across layers.

For the AI industry, this work signals growing emphasis on embodied AI and spatial reasoning—capabilities increasingly demanded by robotics companies and autonomous vehicle developers. The explicit focus on memory mechanisms suggests future MLLM architectures will prioritize persistent spatial understanding over single-frame analysis. This advancement could accelerate practical deployment of autonomous systems in complex environments requiring accurate spatial navigation and object tracking.

Looking forward, the research establishes spatial memory as a measurable, benchmarkable capability for video MLLMs. Continued development in this direction may influence how foundation models are trained and evaluated, particularly those targeting robotics and autonomous applications where spatial coherence across time is non-negotiable.

• →LongSpace-Bench provides a standardized benchmark for evaluating long-horizon spatial memory in video understanding models.
• →The LongSpace framework improves spatial reasoning by incorporating 3D structural cues and layer-aware memory mechanisms.
• →Spatial memory is emerging as a critical capability for autonomous systems, robotics, and embodied AI applications.
• →Explicit memory architectures show measurable improvements over baseline approaches in long-video spatial understanding tasks.
• →Research highlights the gap between current MLLM capabilities and practical requirements for real-world autonomous navigation systems.