So in the new Model Context Protocol Server integration, we get a Stateless Assist option “where the prompt does not contain any state information”. But it’s not available for the built-in voice assistant. This would be great for optimizing llm response time, since a constant prompt could be cached, reducing the time needed to process prompts. And prompts can’t possibly be constant with states inside them. I think the best way to do this is to have constant prompts and have llm retrieve states on demand.