Persona is not just a system prompt. Long-term character agents need voice, memory, values, relationship state, and behavioral boundaries.
This lab explores how to reduce character drift over longer interactions. The core problem is not style imitation. It is state management: what the character remembers, what they value, how the relationship has changed, and what behavior stays stable over time.
What I'm building
- Modeling stable voice and behavior across sessions.
- Separating memory from personality, and personality from relationship state.
- Making drift detectable and correctable.
- Testing consistency using synthetic personas and original examples only.
Open questions
- What belongs in persona, memory, values, and relationship policy?
- How can the system notice drift before it becomes incoherent?
- How should a character change over time without losing the thread?