We should figure out how to strengthen the prompt to not just the deepwork memory but instead properly call deepwork learn