evie-agent · Persistent multimodal agent state machine

This prototype connects the visual and sound assets produced by LoRA & TTS Studio to the Runtime, and maintains coherent personality and cross-modal scheduling through persona state, emotion state, relationship state and user memory. It verifies how the persistent Agent scenario consumes structured state, and does not assume the overall narrative of the three core products.

Clear memory boundaries: Development Infrastructure (vault / Claude Code skills) is only used as a tool and project document when I develop Evie.Do not enter the user-mode memory of Evie Runtime. Evie Runtime separately maintains four types of runtime states: persona state, emotion state, relationship state and user memory, avoiding the mixing of development knowledge base, engineering documents and user relationship memory. The downstream Interaction layer (mainly desktop Local Chat, PWA is a lightweight branch) presents the conversation surface to the user, and the Embodiment layer (ue-realtime) translates the emotional state into a visible body.

Therefore, the most important task at this level is not "whether the answer is good or not", but "whether the persona is stable, whether the decision-making is consistent, and whether the multi-modality is correctly arranged in the long-term existence of cross-tools".

Virtual humans are not "disposable prompts". The mainstream chatbot paradigm is stateless: each conversation starts from scratch with context, personality is forcibly maintained by system prompts, and is broken across tools. This paradigm is fundamentally misaligned with the vision of "lasting digital life".

What evie-agent wants to solve is: treating multiple MLLMs as tool To orchestrate, instead of treating MLLM as a chatbot. The agent itself holds the persona schema + emotion state + cross-day memory, and then dispatches the request to the most appropriate provider according to the scenario (conversation to LLM A, graph generation to SDXL pipeline, expression to emotion model, voice to TTS). Persona must be stable across conversations, and emotion → expression must be controllable.

If this layer is missing, the virtual human vision will be stuck at "looking like the same person, but like a new acquaintance every time we meet" - this is the dead end of all current character AI products.

The whole thing is in the building stage. The skeleton has been formed, but the two main lines of "continuous presence" and "emotional mapping" are still being explored, and there is still some way to go before a truly self-driven agent.