MIT researchers developed SEAL, a framework that lets language models continuously learn new knowledge and tasks.