I haven’t run extensive tests yet, but larger models also get better at memory.