A long-running AI agent can pile up two million tokens of historical input. The bill for a single session: three hundred dollars. Much of that money is not spent on thinking. It is spent on reminding the agent, over and over, of what it did before. One industry analysis puts it plainly: agent amnesia is turning into real invoices.
At this August’s DTCC, the China Database Technology Conference, Tencent Cloud picked this problem up and treated it as a database problem. The conference ran for three days in Beijing, from August 20 to 22, and AI plus database was its dominant theme. Tencent Cloud announced a full slate of AI-oriented products. The flagship newcomer is called TDSQL Nexa, positioned as a unified data entry point for agents. Wang Yicheng, vice president of Tencent Cloud, offered a judgment: when enterprises choose a database in the future, they will look at two things at once, whether it can hold up core business systems, and whether it can hold up agent applications. He called this the new dividing line in database selection.
To understand that dividing line, look at what kind of user an agent is. When people use a database, they know the business first, then look at the tables. They write SQL slowly, fix their mistakes, and run it again. Agents are different. They carry no business knowledge in their heads. Table names, field names, metric definitions mean nothing to them; the database has to explain everything. They are fast and voracious. One task can issue more requests than a person would send in a month. They also have no memory. What was said in the last round is completely forgotten in the next. Above all, no one is watching over their shoulder, and the moment permissions are not clearly drawn, a dangerous operation gets executed just like that.
The four capabilities Wang laid out, multimodal unified support, agent-native architecture, semantics and memory, governance and autonomy, map one to one onto these four mismatches. The released products each take one slice. The semantic layer stores table names, fields and metric definitions, then explains them to the agent. Tencent Cloud says that with this layer, the SQL an agent generates needs almost no manual correction. The memory piece is Agent Memory, open-sourced in mid-May this year. Its stars passed fifteen thousand in eighty days, it topped GitHub Trending more than once, and by the time of the conference it had crossed twenty thousand. In early August it released a team-memory edition, letting multiple agents share experience.
There were plenty of numbers at the launch. Against open-source combinations, Tencent Cloud claims TDSQL Nexa delivers more than ten times the compute performance at over sixty percent lower cost. After plugging in Agent Memory, task success rate rises from sixty percent to eighty, while cost drops by 66.3 percent. All of these numbers come from Tencent Cloud’s own tests. They are fine as a reference, but do not take them as third-party conclusions. There are two numbers I rate more highly. One is the star count, which is developers voting with their feet. The other is that token bill, which belongs to no single company; everyone building agents has to pay it.
Of the four capabilities, the one I watch most closely is memory, for a simple reason. There is no standard there yet. Neither MCP nor A2A, the two protocols that connect agents, defines persistent memory, and industry analysis calls it the only infrastructure layer of the agent era that has not been standardized. Plenty of players have already entered. Overseas there are Mem0, Zep and Letta, and several large model vendors are building memory into their own managed offerings. Tencent Cloud’s route is open source. The repo uses pure local SQLite, has zero cloud dependency, carries the MIT license, and anyone can take it and use it. There is another voice in the community too, arguing that PostgreSQL is the real memory layer. Databases want to swallow memory, and memory wants to grow on top of databases. Which side wins is too early to say.
For DBAs, the change is already visible. When agents connect to databases, someone has to decide which tables they may read, which operations require human confirmation, and how audit trails are kept. When agents need to experiment, database branching appears, spinning up an isolated environment in seconds and rolling back when it breaks. Memory governance will gradually land in DBAs’ hands as well. The hardest part of memory comes after remembering: whether it can be deleted, and whether true can be told from false. Neither has a reliable answer yet. None of this work appears in the operations manuals of three years ago.
The dividing line Wang spoke of may really become one. Database selection used to ask how many transactions per second a system can take, and how backup and recovery work. Going forward, there will be one more question. When the agents come, can you handle them?
Sources
ITHome, report on Tencent Cloud unveiling TDSQL Nexa and multiple AI database capabilities at DTCC, 2026-08-20 Chinaz, report on Tencent Cloud VP Wang Yicheng’s talk on the four core capabilities of AI databases, 2026-08-21 China Daily, “Tencent Cloud Agent Memory breaks 10,000 stars 80 days after open-sourcing, unveils Team Memory”, 2026-08-07 CSDN, industry analysis on agent memory becoming new infrastructure, Tencent topping GitHub and giants racing for the memory layer Ji Yin Jin Morning Report, Issue 83, on-site information from DTCC 2026