RoboMemory is a brain-inspired multi-memory framework for lifelong learning in physical implementation systems. It addresses critical challenges such as continuous learning in real-world environments, multi-module memory latency, capturing task correlations, and mitigating infinite loops in closed-loop planning. Based on cognitive neuroscience, it integrates four core modules: an information preprocessor (thalamus-like), a lifelong implementation memory system (hippocampus-like), a closed-loop planning module (prefrontal cortex-like), and a low-level executor (cerebellum-like), enabling long-term planning and cumulative learning. The lifelong implementation memory system, at the heart of the framework, mitigates the inference speed issues of complex memory frameworks through parallelized updates/retrieval across spatial, temporal, episodic, and semantic submodules. It integrates a dynamic knowledge graph (KG) and a consistent architectural design to enhance memory consistency and scalability. Evaluation results on EmbodiedBench show that RoboMemory achieves a new state-of-the-art (SOTA) performance that is 25% higher on average than the open-source benchmark (Qwen2.5-VL-72B-Ins) and 5% higher than the closed-source state-of-the-art (SOTA) benchmark (Claude3.5-Sonnet). Elimination studies validate key components (criticism, spatial memory, and long-term memory), and real-world deployments demonstrate its lifelong learning capabilities, significantly improving success rates in repetitive tasks. RoboMemory mitigates high-latency challenges through scalability and serves as a baseline for integrating multi-modal memory systems into physical robots.
