Isn't this already the case? Perhaps you mean in a non-transient fashion, i.e. internalizes the in-context learning into the model itself, sort of an ongoing training, that isn't sort of a "hack" like writing notes or adding to a RAG database or whatever.