Phoenix V2 Introduces a Local-First Persistent AI Assistant Architecture

Summary

Phoenix V2 is the companion codebase for Cleverson Santos’s book and paper on persistent AI assistants. It presents a local-first architecture in which memory, identity, and emotional state are stored in SQLite rather than delegated to the language model, allowing them to survive model changes, restarts, and conversation resets. Each input passes through five specialized agents for memory retrieval, planning, action selection, reflection, and personality. The agents share a per-request in-memory blackboard, while the persistent database stores memories, emotional state, and the assistant’s self-model. A Pleasure-Arousal-Dominance engine updates emotion from interaction history, and separate Daydream and Subconscious processes generate reflections and consolidate memories during idle or rest periods. User positive and negative feedback contributes to an internal reinforcement score, while semantic and priority scoring guide memory retrieval. The repository includes a React interface, Express API, SQLite storage, scheduling, tool routing, multi-user profile code, and a Gemini API wrapper; it requires Node.js 18 or later and a Gemini API key. The project is designed to run on modest hardware such as an Intel Core i3 laptop with 8 GB of RAM, but it is explicitly not production-hardened, has no authentication, and is not intended for multi-user deployment as-is. It is released under the MIT License.