Do you think a similar model architecture could include some sort of memory, so a follow-up prompt of "Go back to the water foutain" would somehow work?
A in-model memory approach is probably still deep research but maybe a Rag-like pipeline could work in some instances