LLMs with self-editing memory for unbounded context