I built an AI companion that implements *state stability and persistent persona evolution*—the next step beyond simple RAG or chat history. The system uses Python/Flask and Gemini to manage a *multi-dimensional personality state* (Tsundere, Yandere, etc.) that evolves over time. ### Core Technical Challenge & Solution: 1. *State Stability (Hysteresis)*: We prevent the persona from "flipping" instantly based on one message by introducing *inertia* (requiring 2x the score difference to change an established personality). 2. *Affection Scaling*: The emotional depth (Affection Level) scales infinitely, leading to complex state shifts over long interactions. 3. *Efficiency*: All logic runs within a single, cost-optimized API call. ### I'm looking for technical critique on: * *Architecture*: Are hard margins (10-point difference) for stability the best approach, or should this be solved with a purely ML-driven method? * *Safety & Alignment*: How to maintain the provocative nature of personas (like Yandere) while adhering to strict safety boundaries. GitHub Repo: [https://github.com/EMMA019/Evo-chat] |