It sounds like this is a big deal. I thought most AI models were built this way already though, as operating systems for LLM prompt generation. This seems like an iteration of that that includes an ability to store certain data it's referring to frequently as "facts", and let other data degrade over time to allow it to learn new things. Is that accurate?