Dynamic memory architecture for use with large language model(s)
Topics: AI Mode, LLMO / GEO, Personalization, Probably in use, Search Intent, Search Query Processing
This Google patent describes a system for an automated assistant that extracts and stores “inferences” (assumed facts or preferences) from a user’s natural language input. Unlike standard systems that either forget information quickly or store it rigidly, this architecture uses a dual-model approach: one model generates an initial inference with a confidence score, and a second model continuously updates that score as new data arrives. The system then decides whether to include these refined inferences in future prompts to a Large Language Model (LLM) to ensure responses are personalized but based only on reliable, up-to-date information.
