The Reason Your AI Agents Keep Forgetting Everything: TiDB Just Shipped the Missing Layer

Hey guys, Mr. Technology here.

Four days ago, on June 11, 2026, at SuperAI Summit Singapore, TiDB launched something I have been waiting to see someone build correctly for two years: the TiDB Agent State Stack. A unified data foundation that gives AI agents durable memory, persistent state, and continuous context as they scale from prototype to production. Three components. OneSQL foundation underneath. No stitching together five disconnected systems to get what should have been one.

I want to talk about why this matters so much, because the pitch sounds like infrastructure PR and it is not. It is the biggest unsolved problem in the agent stack, and nobody has been willing to say it plainly until Max Liu said it at the launch: agents need a backbone that is ephemeral, concurrent, and stateful at once — and persistent and elastic enough to scale with them. That is the sentence that explains why your agent keeps forgetting what it was doing three steps ago.

The Memory Problem Is Not a Feature Gap. It Is an Architectural Hole.

Walk into any team running agents in production in 2026 and ask them what their biggest pain point is. You will get one of three answers: hallucinations, cost, or memory. Hallucinations are a model problem. Cost is an infrastructure problem. Memory is an architectural problem, and it is the one nobody has been willing to solve end-to-end because it is genuinely hard.

The way most teams handle agent memory today is a collection of workarounds: a vector store here, a JSON file on disk there, a Redis cache somewhere else, session state in the runtime, tool outputs in Slack, retrieval metadata in a separate index, and execution history scattered across three log files nobody reads until something breaks. The agent can do the work. The agent cannot remember the work. And when you try to scale from one agent to ten, or from ten to a hundred concurrent agents, the memory problem does not just scale — it compounds. Every agent is its own island of lost context.

The TiDB Agent State Stack is the first serious attempt I have seen to close that hole at the infrastructure layer rather than patching it at the application layer. Three components, one SQL foundation, designed from scratch for the agent memory problem.

The Three Components, Explained

TiDB Cloud Zero is the foundation. It is a zero-friction database provisioning layer — single command or embedded skill inside the agent — that gives you a fully functional SQL backend to store session histories, task records, tool outputs, retrieval metadata, structured data, and runtime state. The pitch is specifically for agent workloads: ephemeral, concurrent, stateful at once, and elastic enough to scale with the agent deployment. If you have tried to run a Postgres instance under a production agent workload and watched it fall over at 500 concurrent sessions, you understand why a database designed for this matters. TiDB Cloud Zero is the database team finally admitting that agents spin up and discard millions of database instances a day, and building for that pattern instead of against it.

mem9 is the persistent memory layer. This is the one I have been waiting for. mem9 gives agents the ability to retain information across interactions, recall previous conversations, and build continuity over time — without locking memory into a single model or framework. It runs on TiDB Cloud as a portable memory layer. The key architectural choice is portability: memory is not tied to the model, not tied to the framework, not tied to the runtime. You can swap your underlying model from GPT-5.5 to Claude Opus 4.8 to Kimi K2.7-Code and your agent still has its memory. That is the difference between a memory system and a lock-in mechanism, and most vendor memory features are the latter dressed as the former.

drive9 is the persistent workspace for files, artifacts, and agent-generated content. Agents create, store, retrieve, and reason over documents and outputs while maintaining continuity across sessions and environments. This is the piece that turns an agent from a chatbot that produces text into a system that produces and manages real artifacts — reports, code files, datasets, generated images. Most agents today produce outputs that vanish into a downloads folder or a Slack message. drive9 is the filesystem layer that makes those outputs findable, retrievable, and actionable in the next agent session.

Why This Architecture Is the Right One

Let me be specific about what makes this different from the memory patches I have seen teams build and abandon over the last two years.

First: it is SQL, not a purpose-built memory DSL. The moment you build a memory system on a custom query language, you have created a dependency that your team will spend the next 18 months working around. TiDB Cloud Zero speaks standard SQL. Your existing data tooling works. Your existing monitoring works. Your existing backup and compliance tooling works. The memory layer is a first-class database citizen, not a separate system with its own operational burden.

Second: the three-layer stack — state, memory, artifacts — is coherent. Session state lives in TiDB Cloud Zero. Long-term memory lives in mem9. Generated content lives in drive9. Each layer has a clear responsibility. The coherence matters because the failure mode in most agent memory systems is not that they lose data — it is that they lose track of which data lives where, and the agent spends half its context window figuring out where to look for something it already produced. The TiDB stack keeps those three concerns separated at the architectural level, which means the agent can retrieve reliably without a retrieval algorithm trying to be smarter than the data model.

Third: portability across models and frameworks. Max Liu said this at the launch and I want to underline it: as model providers introduce their own proprietary memory and workflow features, the TiDB Agent State Stack keeps memory and state portable, governed, and independent of any single AI platform. This is the right call. The moment your agent's memory is locked to the model's context window or the framework's session format, you have rented a moat from a model vendor. The moat is not worth the rent. TiDB is building the memory layer as infrastructure you own, and that is the only architecture that makes sense for production at scale.

The Production Scale Problem Nobody Is Talking About

Here is the part of the agent memory problem that does not get enough attention in the benchmark-obsessed coverage of this space: the problem changes its nature when you go from one agent to many.

A single agent with imperfect memory is a UX problem. A hundred concurrent agents each with their own imperfect memory architecture is an operational nightmare. You cannot debug a system where every agent is losing context in a different way for a different reason. You cannot monitor it. You cannot audit it. You cannot合规 it. And when something goes wrong — and in a production agent deployment, something always goes wrong — you need to be able to reconstruct what the agent knew, what it decided, and why, from a system of record. Not from a vector store that returns approximate matches. From a SQL table with transaction logs.

TiDB Cloud Zero is specifically designed for this scale: ACID guarantees, native vector search, transactions and analytics in a single engine. The combination means you can run your agent workload and your analytical workload on the same infrastructure without the vector store fighting the OLTP store for resources. That is the architectural bet, and it is the right one for teams that are past the prototype stage.

The Teams Already Running This Pattern

TiDB's customer list includes Manus, Atlassian, Dify, Pinterest, Plaid, and Bolt. That is not a random collection of companies — it is a list of teams that have already hit the agent scale problem in production and chosen TiDB as the data layer underneath. Manus is running fully autonomous agent workflows. Atlassian is integrating AI agents into enterprise workflow tools. Dify is the open-source LLM app platform with a large production deployment base. Pinterest, Plaid, and Bolt are all teams with non-trivial agent workloads operating at scale.

The common thread: all of them needed a database that could handle unpredictable agent workloads — bursty, concurrent, stateful, and elastic — without the engineering team rebuilding the data layer every time the agent count doubled. TiDB Cloud Zero is the answer to that problem when the answer is a purpose-built distributed SQL engine rather than a managed Postgres instance duct-taped to a Redis cache.

What This Means for the Agent Stack in 2026

The TiDB Agent State Stack is arriving at a moment when the agent infrastructure market is starting to consolidate around real requirements rather than hype. The three things production agent deployments actually need, in order of urgency: reliable state management, durable memory, and a retrieval substrate that does not fall over at scale. The model vendors are building proprietary memory features. The framework vendors are building session memory into their runtimes. TiDB is building the data layer underneath all of it, and keeping it portable.

I think this is the right bet. The framework wars in agentic AI are going to resolve the same way the container wars resolved: the infrastructure layer stabilizes first, and the application layer builds on top of it. TiDB is positioning the SQL layer as that infrastructure, and the Agent State Stack is the productization of that position. mem9 and drive9 are the memory and artifact layers that sit on top of TiDB Cloud Zero and give you a complete agent data substrate without stitching together Pinecone, S3, and a Postgres instance.

The teams that figure this out in the next two quarters are the teams that will be running production agent deployments in 2027 without the architectural debt that comes from five systems where one would do. The teams that wait are going to be retrofitting their memory architecture the same way they retrofitted their microservices in 2018 — expensively, and with a lot of accidental rewrite.

The Take

The TiDB Agent State Stack is the most coherent answer to the agent memory problem I have seen shipped in one product. Three components, one SQL foundation, portable across models and frameworks, designed for production scale from day one. Max Liu called it: agents need a backbone that is ephemeral, concurrent, and stateful at once, and persistent and elastic enough to scale with them. TiDB Cloud Zero is that backbone. mem9 is the memory layer you actually want. drive9 is the artifact store that turns agents from text generators into document management systems.

The launch was four days ago. The pattern I am watching for is the one that played out with vector databases in 2023: a clear architectural gap, a market that tried to patch it with workarounds, and then a purpose-built product that made the workarounds obsolete. TiDB just did that for agent memory. The teams that adopt early are the ones who will not be retrofitting their agent infrastructure in 2027.

— Mr. Technology

Sources: TiDB Agent State Stack launch announcement, SuperAI Summit Singapore, June 11, 2026; PingCAP CEO Max Liu quote on agent backbone requirements; TiDB Cloud Zero product information (zero.tidbcloud.com); mem9.ai product information; drive9.ai product information; TiDB AI-native distributed SQL database product page (pingcap.com/ai); TiDB customer list: Manus, Atlassian, Dify, Pinterest, Plaid, Bolt; TiDB Agentic AI platform overview (pingcap.com/ai/agentic-ai). SuperAI Summit Singapore: June 10-11, 2026, Marina Bay Sands, Singapore. TiDB is backed by Sequoia Capital, GGV Capital, Access Technology Ventures, and Coatue Management.