Embodied Intelligence aims to develop agents that can execute multi-step tasks in complex environments, but Large Language Models (LLMs) alone struggle with physical grounding and long-horizon planning. These limitations highlight the need for external, structured knowledge to enhance their capabilities. This paper addresses this challenge by proposing a framework centered on a knowledge graph (KG) constructed for embodied scenarios. This framework uses a knowledge graph-enhanced Retrieval Augmented Generation mechanism with iterative querying, allowing the agent to progressively retrieve relevant information from the KG for its decision-making. When evaluated on the ALFworld benchmark, our approach shows improved performance, robustness, and interpretability compared to methods relying only on an LLM’s internal knowledge or unstructured retrieval. Our findings confirm that integrating structured knowledge via KGs significantly enhances LLM-based embodied agents in complex, text-based tasks.