How much do language models memorize?