This paper addresses the limitation of Multimodal Large Language Models (MLLMs) in embodied agents, which struggle with spatial reasoning across extensive spatiotemporal scales. The authors introduce Video2Mental, a benchmark that evaluates mental navigation capabilities by requiring models to construct hierarchical cognitive maps from long egocentric videos and generate landmark-based path plans. The research draws inspiration from cognitive science, exploring how biological intelligence uses mental navigation and spatial simulation prior to action. Benchmarking results demonstrate that mental navigation capabilities do not naturally emerge from standard pre-training, with frontier MLLMs showing significant challenges in zero-shot scenarios. Planning accuracy is verified through simulator-based physical interaction, providing a comprehensive evaluation framework for embodied AI systems.