How can I let an AI assistant learn from user corrections while keeping context lean?