I feel like we’re working on one of the hardest engineering problems at friend. creating, updating, and querying “memories” while chatting with LLMs already barely works, but since friend is always listening, we have so much more data. We also need extremely accurate diarized transcripts as a ground truth, something that also barely works, given that ASR tech is usually used for YouTube videos, not thousands of hours of day to day life. very quickly we create extremely complex “knowledge graphs” of each individual. even a 30 minute conversation alone in a vacuum tells more about you than 20 years of Googling history. *how* you say things matters a lot. however the real challenge is creating a companion that begins as an unformed ball of clay at onset that very quickly evolves into a deeply nuanced personality that fits each user specifically. Ultimately this creates a companion personalization akin to the best social chameleon with whom you've ever interacted. (h/t Jordan Cooper) If you’ve got ideas on how to build such a robust memory system, my DMs are open