Stop Fine-Tuning to Fix Retrieval Problems — Anant Srivastava

This post introduces a framework for managing knowledge in AI systems, differentiating the roles of prompts, memory, and weights. It argues that prompt engineering is for small, stable behaviors; memory is for current, large, citable, or access-controlled knowledge; and weights are for reflexes that have stopped changing. The talk highlights how architectural decisions about knowledge placement are often made accidentally through product work and presents a diagnostic for choosing the right tool. It also proposes a circulating architecture where agents improve by performing their jobs, suggesting a dynamic and evolving knowledge system.