What Spatial Memory Must Store: Occlusion as the Test for Language-Agent Memory

The paper introduces the concept of a "memory palace" system for language agents that integrates spatial geometry into memory recall, demonstrating that geometry-based weighting significantly improves recall accuracy over traditional linear blends. Key results include a mean Delta-Hit@5 improvement of +0.3208 (p<10^-15) for geometry-led recall, and a DDA method achieving a visibility accuracy of 0.982 in occlusion scenarios, compared to 0.000 for conventional methods. This research underscores the necessity of incorporating geometric principles into memory systems, which is crucial for enhancing the performance of AI agents in spatially aware tasks.