Half my feed for the past few days has been realizing the same thing: We gave AI PhD-level reasoning, and it still forgets what you told it yesterday. There's a paper called "The Price of Meaning" that basically puts math behind a problem a lot of builders are running into. When memory is built on embeddings, the system retrieves what is semantically close. Not necessarily what is true. That distinction matters a lot. We learned this the expensive way while building @cfosilvia. For a normal AI app, a fuzzy memory might be annoying. For a product people trust with their finances, fuzzy memory is a real problem. The thing that held up best in production was almost boring: Store exact facts. Preserve history. Separate what the user said from what the model inferred. Let the model navigate memory instead of pretending similarity equals truth. Reasoning gets most of the attention in AI. But memory is what determines whether an agent actually gets better over time, or just feels impressive the first few times you use it.