Phoenix V2: persistent memory, emotional state, self-model for AI (MIT) (github.com)

MIT has introduced Phoenix V2, a groundbreaking AI assistant designed for persistent cognitive functioning, which emphasizes local memory management and emotional awareness. Unlike traditional AI models that rely on cloud-based large language models (LLMs) for memory retention, Phoenix V2 stores user interactions, emotional states, and personal identity directly in a local SQLite database. This architecture enables the assistant to maintain continuity across sessions and withstand model updates, providing a more personalized and consistent user experience. The significance of Phoenix V2 lies in its innovative approach to addressing the well-known amnesia problem in AI systems by utilizing a multi-agent pipeline that includes specialized processes for memory retrieval, planning, action, and reflection. An emotion engine based on the PAD model tracks user's emotional states continuously, while mechanisms like the Daydream and Subconscious Engines facilitate idle reflections and memory consolidation, respectively. This comprehensive design is detailed in Cleverson Santos' book, which accompanies the codebase, making it accessible for developers interested in building persistent AI systems on standard consumer hardware.