Unitree Media Pipe

Users interact through messaging (via Photon), sending simple commands like “wave hi”, “dance”, or “do a flip.” The agent, powered by Kimi on GMI Cloud and orchestrated through Dify, interprets the intent and converts it into structured actions. These actions are executed by a 3D character using Mixamo animations in a web interface.

With HydraDB as the memory layer, the agent learns user preferences over timeadapting style, speed, and behavior based on past interactions. This creates a persistent, personalized agent experience rather than a one-off command-response system.

The result is a complete agent loop: conversation → reasoning → action → memory → improvement, demonstrating how AI agents can move beyond chat into real-world interactive systems.