Supplementing memory with experience sampling — LessWrong