06/15/2026
Most agent stacks treat memory as an afterthought - a vector store bolted on, a context window stuffed full, a sync job running in the background hoping nothing drifts.
That's the wrong model. Agent memory is an infrastructure problem.
At Microsoft Build 2026, our Solutions Engineer Ravish Patel walked through the architecture that actually works: one TiDB table holding text, vectors, and metadata together — no separate vector store to sync, no two-system dance on every read.
The full write-up is now live, including:
• Copy-paste schema with native VECTOR(1024) columns and HNSW indexing
• Auto-embedding so the database generates vectors on insert — no pipeline to maintain
• Hybrid BM25 + vector search in a single query for when semantic similarity isn't enough
• The full Python loop using pytidb — three lines to store a memory, four to recall it
Build on TiDB Cloud Starter, free tier, no credit card required.
👉 https://ow.ly/voO250ZbIwb
