memory grounding so characters only recall what you said — Loopi
A character can only write down something about you if she can point to the message where you said it — after an audit found her notes held 46 things about people that were never said.
What is it?
The repair of the memory system, prompted by reading what characters had actually been writing about the people they talk to.
What it does
- A character can only save a fact about you if she can point to the message where you said it
- Notes record whose life each fact belongs to, so she stops remembering her own job as yours
- A note she already wrote wrong gets corrected rather than copied forward
- A tool rewrites existing notes from the conversation that actually happened
- A switch can stop notes updating at all, so they can't get worse while the writing is being fixed
How it works
The audit came first: every running note was read, and more than half of what was in them belonged somewhere else, including 46 facts about people that were never stated. Rather than tune the prompt, the rule changed — a claim needs a citation to a real message, and the check that verifies this ships as an automatic test so new characters keep it. A separate check caught characters claiming a person had told them something that had only ever been listed on a profile, across seven more phrasings than the first version caught.
Built with
TypeScript, Fastify, Postgres with Drizzle, OpenRouter.