Core Viewpoint - The article discusses the advancements in embodied intelligence, specifically focusing on the 3DLLM-MEM model and the 3DMEM-BENCH benchmark, which enable AI to build, maintain, and utilize long-term memory in complex 3D environments, addressing the limitations of existing large language models (LLMs) in spatial-temporal memory management [3][10]. Group 1: Challenges in 3D Environments - Existing LLMs excel in text understanding but struggle in dynamic 3D environments due to their reliance on sparse or object-centric representations, which fail to capture complex geometric relationships crucial for task success [5][6]. - The lack of a dynamic updating mechanism in current models leads to outdated memories, making it difficult to distinguish between old memories and new states [5][6]. - In multi-room tasks, models often fail to associate observations across different times and spaces, resulting in critical information being forgotten [8] [10]. Group 2: Breakthroughs with 3DLLM-MEM and 3DMEM-BENCH - The 3DMEM-BENCH benchmark is the first to evaluate long-term memory in 3D environments, featuring over 26,000 trajectories and 1,860 embodied tasks across 182 3D scenes [11][13]. - The benchmark includes multi-dimensional assessments and difficulty levels ranging from simple to challenging tasks, testing the model's generalization capabilities [12][13]. - The 3DLLM-MEM model introduces a dual-memory architecture that integrates working memory and episodic memory, allowing for selective retrieval of relevant features while avoiding memory overload [16][19]. Group 3: Performance Validation - The 3DLLM-MEM model significantly outperforms baseline models, achieving a success rate of 27.8% in the most challenging "wild difficulty tasks," compared to only 5% for recent memory models [21][23]. - In spatial reasoning tasks, the model achieves over 60% accuracy, while traditional 3D-LLMs struggle with less than 10% accuracy due to contextual limitations [24]. - The model's dynamic fusion mechanism reduces computational costs by processing only task-relevant memory segments, maintaining high inference accuracy [25].