I’m testing different ways to maintain a “long-term” memory in conversational chatbots. From your experience, which methods have given you the best results? Eg. RAG/Vector Database + the last X messages in the prompt, …