Phoenix V2: A Cognitive Architecture for Persistent, Emotionally-Aware AI Assistants on Consumer Hardware

Large Language Models (LLMs) are stateless by design: each session begins without access to prior interactions, user preferences, or accumulated relational context. This architectural property—which we term the amnesia problem—prevents LLMs from functioning as persistent collaborative partners. We present Phoenix V2, a cognitive architecture that decouples persistent intelligence from the reasoning engine, treating the LLM as an interchangeable consultant while keeping memory, affective state, and identity in a local, user-controlled SQLite file. Phoenix V2 integrates five subsystems: (1) an episodic memory store with semantic retrieval over embedding vectors; (2) a multi-agent cognitive pipeline coordinated through a Blackboard; (3) an affect model in a Pleasure–Arousal–Dominance (PAD) space that persists across sessions; (4) a SubconsciousEngine that performs background consolidation, reflection, and self-model updates during idle cycles; and (5) an incremental learning layer that extracts behavioral patterns without model retraining. The system is implemented in TypeScript on Node.js, uses no local GPU, and was developed and runs on a 2008-era notebook (Intel T9300, 6 GB RAM). We describe the design principles and the shipped implementation, position it against Mem0, MemGPT/Letta, Zep, and Generative Agents, and state its limitations—notably the absence of quantitative evaluation and the current reliance on a remote embedding API. The companion code is released under the MIT License at https://github.com/cleversonbrsantos-art/Phoenix.