Self-Learning Character

Preamble

Self-Learning Character is the proof branch where a fictional person acts, reflects, proposes a small self-model change, and has to prove the change through later behavior.

---

A Prompt Cannot Develop

Static character prompts can imitate depth. They can also repeat the same wound, virtue, fear, and voice forever. Development requires pressure from events and a record of what those events changed.

episode -> behavior -> reflection -> self-model mutation -> held-out scene -> keep or reject

The character’s self-model stores identity, voice, values, fears, goals, relationship map, known contradictions, and behavioral heuristics. Episodic memory stays separate so one dramatic scene cannot rewrite the whole identity by accident.

The Mutation Has To Behave

A proposed change might say the character has become less trusting after betrayal. The next test is behavioral: what do they do when trust would be useful, costly, or humiliating?

The evaluator checks coherence, richness, drift, generic prose, and memory pollution. A rejected mutation still shows what the character tried to become and what the system refused.

The Demo Waits

The public chatbot comes later. This branch first needs the private machinery: episodes, versioned state, accepted mutations, rejected mutations, and watchtower output that shows whether the character became richer or only better at describing itself.