RAG isn't memory. It's Ctrl+F with embeddings.

This article critiques the common use of Retrieval-Augmented Generation (RAG) as "memory" in AI agents. The author argues that RAG, based on vector search, is fundamentally just search, not true memory, leading to issues like agents retrieving outdated or irrelevant information. Three critical flaws are identified: chunks lacking contextual awareness, the absence of structure beyond cosine similarity, and the lack of time as a first-class concept. The author dismisses current "memory frameworks" as insufficient, as they lack core properties like a formal "I don't know" mechanism. The article also challenges the trend of relying on extremely long contexts, pointing out cost, recall issues, and lack of persistence. The author advocates for replacing the word "memory" with "retrieval", utilizing temporal validation, and making "I don't know" a possible outcome. The piece concludes by emphasizing that true memory is possible and hinting at part 2, which will delve into seven architectural principles.