Your agent's memory is a vector store. Ask it "how many" and watch it fall over.

Short summary

Vector stores solve semantic recall for AI agents but fail at analytical queries like counts, groupings, and rankings because they only do nearest-neighbor search with no SQL primitives. When an agent needs to answer 'how many users asked about pricing,' the top-k retrieval is structurally incapable of producing a correct count. The solution is to compose a vector store for recall with a relational store for analysis, picking the right machine per question shape.

• •Vector stores lack COUNT, GROUP BY, and JOIN — analytical queries return confident wrong numbers
• •LLMs doing arithmetic over retrieved samples hallucinate rather than compute
• •Compose vector store (recall) with relational store (analysis) for complete agent memory

Generated with AI, which can make mistakes.