language models typically need to be trained or finetuned in order to acquire new knowledge, which involves updating their weights. We instead envision →