AI can now learn from your preferences without storing your data or retraining itself

A technique called AdaptFuse lets language models update what they know about a user's preferences across multiple conversations without fine-tuning on that user's data. Instead of the model learning and storing personal information, a separate symbolic system keeps track of what the model is inferring about your preferences, and they work together to make better predictions.

Privacy-sensitive applications like healthcare and finance have always faced a tradeoff: either the AI learns from user interactions and stores personal data, or it stays dumb about what you actually need. This removes that tradeoff by separating learning from storage — the inference happens, but the model itself never touches your data and never changes. Companies building personalized systems can now offer better recommendations without building a database of user behavior, which matters because they can't be breached, sold, or subpoenaed.