Francesca sharpens companion personality, memory

The model is only part of the story here; the moat is the orchestration layer that filters generic replies, controls memory, and catches self-contradictions.

• –Generating three candidates and ranking them for crutch phrases is a practical way to keep the persona from collapsing into generic therapist mode.
• –The opener experiment is the most interesting product insight: grounded specifics seem to retain users better than vague psychoanalysis.
• –Proportional memory with category caps is the right compromise for companion apps; unlimited memory tends to fossilize one user’s quirks into the whole persona.
• –Self-fact guards are underrated, because tiny mirroring mistakes feel much bigger in intimate chat than they do in ordinary chatbot use.
• –Voice cloning plus local inference makes the product feel embodied, but it also raises the consistency bar across every layer.