ELM is an embodied language model designed for understanding long-horizon driving scenarios in space and time. Unlike traditional vision-language models (VLMs), which are limited to 2D domain and lack spatial awareness and long-term forecasting capabilities, ELM incorporates space-aware pre-training to enhance spatial localization and time-aware token selection for accurate temporal cues. ELM is evaluated on a new benchmark with ten tasks, including description, localization, memorization, and forecasting. The model outperforms previous state-of-the-art approaches in all aspects. ELM is trained on a diverse dataset including nuScenes, Waymo, YouTube, and Ego4D, enabling it to handle complex driving scenarios. The model's performance is validated on various tasks, demonstrating its effectiveness in long-term event reasoning and forecasting. ELM's architecture includes pre-training on open-world data and fine-tuning on diverse tasks. The model's time-aware token selection module enables efficient retrieval of relevant information from long-term memory. ELM's results show significant improvements over other models, including LLaMA-Adapter V2, LLaVA, Otter, and VideoChat. The model's ability to handle long-horizon scenarios and its robust performance in various tasks highlight its potential for autonomous driving applications.ELM is an embodied language model designed for understanding long-horizon driving scenarios in space and time. Unlike traditional vision-language models (VLMs), which are limited to 2D domain and lack spatial awareness and long-term forecasting capabilities, ELM incorporates space-aware pre-training to enhance spatial localization and time-aware token selection for accurate temporal cues. ELM is evaluated on a new benchmark with ten tasks, including description, localization, memorization, and forecasting. The model outperforms previous state-of-the-art approaches in all aspects. ELM is trained on a diverse dataset including nuScenes, Waymo, YouTube, and Ego4D, enabling it to handle complex driving scenarios. The model's performance is validated on various tasks, demonstrating its effectiveness in long-term event reasoning and forecasting. ELM's architecture includes pre-training on open-world data and fine-tuning on diverse tasks. The model's time-aware token selection module enables efficient retrieval of relevant information from long-term memory. ELM's results show significant improvements over other models, including LLaMA-Adapter V2, LLaVA, Otter, and VideoChat. The model's ability to handle long-horizon scenarios and its robust performance in various tasks highlight its potential for autonomous driving applications.