I've done 750+ interviews on The Ravit Show. Everyone asks about models. Frameworks. Platforms. Almost nobody asks the question that actually decides whether an AI agent works: where does the memory live? So I sat down with Ed Huang, Co-Founder and CTO of TiDB, powered by PingCAP in Mountain View, and we went deep on the layer everyone is ignoring. A few things from this conversation that stuck with me: → "Memory is the surface, state is the system." Your agent remembering your name is memory. Your agent forgetting what it already tried three steps ago? That's a state failure — and it looks like a dumb agent, even on a frontier model. → Teams stitch together relational + vector + cache + sync pipelines. It works in the demo. It dies at scale. Ed breaks down why collapsing it into one distributed SQL engine matters beyond just "fewer parts." → The laptop-return story: an agent confidently answering from a 2023 policy doc. Better embeddings can't fix it. Ed explains why the retrieval accuracy gap is an architecture problem, not a model problem. → Manus runs 1.2M database clusters — and 99% were created by agents, not engineers. What breaks when your database's "user" is an agent instead of a DBA? Almost everything you assumed. → And the big one: three years out, when everyone has access to the same models, what do AI products actually compete on? Ed's answer — the most reliable memory wins, not the biggest model. If you're building agents, this is the conversation about the layer underneath everything else. Thank you, Ed, for the depth and honesty in this one. #data #ai #agenticai #TiDB #PingCAP #theravitshow
I've done 750+ interviews on The Ravit Show. Everyone asks about models. Frameworks. Platforms. Almost nobody asks the question that actually decides whether an AI agent works: where does the memory live? So I sat down with Ed Huang, Co-Founder and CTO of TiDB, powered by PingCAP in Mountain View, and we went deep on the layer everyone is ignoring.
A few things from this conversation that stuck with me:
→ "Memory is the surface, state is the system." Your agent remembering your name is memory. Your agent forgetting what it already tried three steps ago? That's a state failure — and it looks like a dumb agent, even on a frontier model.
→ Teams stitch together relational + vector + cache + sync pipelines. It works in the demo. It dies at scale. Ed breaks down why collapsing it into one distributed SQL engine matters beyond just "fewer parts."
→ The laptop-return story: an agent confidently answering from a 2023 policy doc. Better embeddings can't fix it. Ed explains why the retrieval accuracy gap is an architecture problem, not a model problem.
→ Manus runs 1.2M database clusters — and 99% were created by agents, not engineers. What breaks when your database's "user" is an agent instead of a DBA? Almost everything you assumed.
→ And the big one: three years out, when everyone has access to the same models, what do AI products actually compete on? Ed's answer — the most reliable memory wins, not the biggest model.
If you're building agents, this is the conversation about the layer underneath everything else.
Thank you, Ed, for the depth and honesty in this one.
#data #ai #agenticai #TiDB #PingCAP #theravitshow