AlphaAvatar: a self-hostable realtime full-multimodal personal AI assistant runtime

Thanks for the thoughtful feedback. I think you captured a direction I’ve been converging toward as well. While AlphaAvatar currently includes visible components such as voice, avatar, channels, RAG, MCP tools, and memory, I’ve been increasingly viewing those as adapters around a reusable runtime rather than the runtime itself. Some of the work I’m doing recently is moving in exactly that direction: * A provider abstraction layer to decouple runtime logic from specific LLM, ASR, TTS, memory, retrieval, and tool implementations. * Multi-user support, where identities, memories, sessions, and relationships are managed independently instead of assuming a single user per runtime. * A more structured conversation/session state model rather than treating everything as chat history. * Exploring multimodal memory, where memory is not limited to text but can eventually incorporate voice, visual observations, identities, interactions, and other modalities. I also strongly agree that memory is fundamentally a policy problem, not just a storage problem. Questions like what should be remembered, when it should be committed, how interruptions affect memory, and how durable memory differs from temporary interaction state are becoming increasingly important as interactions move from chat to realtime multimodal experiences. One thing I find particularly interesting in your feedback is the distinction between generated output and delivered output. In realtime voice interactions those are often not the same thing, and that has implications for memory, replay, evaluation, and user experience. Longer term, my hope is that AlphaAvatar can evolve into something closer to a reusable character/conversation runtime layer that different avatar systems, channels, tools, and multimodal frontends can build on top of, rather than being tied to a single application experience. Really appreciate the detailed thoughts and roadmap suggestions.