Accumulating Context Changes the Beliefs of Language Models