Mind-Map Agent: Enhancing Cooperative Task Planning through Communication Alignment with Large Language Models

Embodied agents that collaborate with humans through natural language have become an active area of research, offering flexibility in cooperative planning and execution. Debate-based approaches often depend on repeated consensus procedures, which can increase dialogue frequency and risk over-communication. At the same time, LLMs are prone to hallucination during dialogue processing, sometimes causing confabulation and reducing consistency in long-term strategies. We introduce the Mind-Map Agent, an approach that guides reasoning with explicit cooperative strategies while maintaining structured long-term memory to disentangle dialogue, task state, and planning context. The generated Mind-Maps support coherent long-horizon planning, reduce redundant dialogue, and enhance interpretability in multi-agent interaction. Evaluations on Communicative Watch-and-Help and ThreeDWorld Multi-Agent Transport indicate that the Mind-Map Agent achieves more stable efficiency compared to classical planners and LLM agents across different model scales and environments. Our results suggest that Mind-Map reasoning enables cooperative agents to accomplish tasks with fewer conversations while sustaining effective collaboration.