Vector stores can't count: why AI agents need relational databases for analytics

AI agents commonly use vector stores to handle memory by embedding facts and retrieving the most similar results for a given query. While this approach works well for fuzzy recall tasks like finding related conversations, it fundamentally fails when asked analytical questions such as counting, grouping, or ranking stored data. Vector stores have no equivalent of SQL's COUNT or GROUP BY, meaning an agent asked 'how many users mentioned pricing this month?' will silently return a confident but incorrect answer derived from only a small sample of records. The recommended fix is a two-layer architecture: a vector store for similarity-based recall and a relational database with a proper query planner for any aggregation or analytical work. The article argues that neither system substitutes for the other, and that scaling up the underlying language model does not bridge the gap between the two fundamentally different retrieval mechanisms.