Multi-Model Collaborative Persona Architecture: A Paradigm for Overcoming Context Window Limitations in AI Companion Systems

This technical report proposes a novel architecture for AI companion systems in which multiple specialized small language models (3B–7B), each trained on domain-specific data via CPT, SFT, or DPO, serve as cognitive subsystems behind a single large language model interface. The large model retains identity and decision-making; the small models handle memory, information synthesis, preference alignment, and persona maintenance. This structurally eliminates context window accumulation and enables continuous persona evolution through automated data feedback loops.