The paper introduces the Retrieval-Augmented Embodied Agent (RAEA), a novel framework designed to enhance embodied agents in complex and uncertain environments. RAEA leverages an external policy memory bank containing a wealth of multi-embodiment data to improve the agents' performance. The system includes a policy retriever and a policy generator. The policy retriever processes multi-modal inputs (text, audio, images, videos, and point clouds) to retrieve relevant policies from the memory bank, while the policy generator integrates these policies into the learning process, enabling the agents to respond effectively to tasks. Extensive testing in both simulated and real-world scenarios demonstrates that RAEA outperforms traditional methods, showcasing its superior performance and practicality in robotic applications. The paper also discusses related works and provides a detailed methodology, including the design of the policy retriever and generator, and evaluates the effectiveness of RAEA through various benchmarks and real-world experiments.The paper introduces the Retrieval-Augmented Embodied Agent (RAEA), a novel framework designed to enhance embodied agents in complex and uncertain environments. RAEA leverages an external policy memory bank containing a wealth of multi-embodiment data to improve the agents' performance. The system includes a policy retriever and a policy generator. The policy retriever processes multi-modal inputs (text, audio, images, videos, and point clouds) to retrieve relevant policies from the memory bank, while the policy generator integrates these policies into the learning process, enabling the agents to respond effectively to tasks. Extensive testing in both simulated and real-world scenarios demonstrates that RAEA outperforms traditional methods, showcasing its superior performance and practicality in robotic applications. The paper also discusses related works and provides a detailed methodology, including the design of the policy retriever and generator, and evaluates the effectiveness of RAEA through various benchmarks and real-world experiments.