On July 15, 2026, millions of users in China will lose the AI agents they built inside Doubao and Qwen. Not because the models failed. Because persistent memory is structurally incompatible with how the regulation is written - and neither ByteDance nor Alibaba could retrofit their way out.
China's first dedicated regulatory framework for AI services that simulate human personality, the Interim Measures for the Administration of AI Anthropomorphic Interactive Services, takes effect on July 15, 2026.
Both companies are pulling their agent features entirely rather than rebuilding them from the ground up, and users who do not export their data before the deadline risk losing it permanently.
Alibaba's Qwen disabled user-created agents on July 10 and will fully shut down agent functions by July 15.
Qwen confirmed permanent deletion of agent data with no migration plans.
The shutdown is real news. But the more durable story is the technical reason it happened.
Why persistent memory and these rules don't fit together
The Interim Measures require anti-addiction systems and two-hour break reminders that clash with how persistent-memory companions are built.
The measures require anti-addiction systems, identity checks for users under 14, and an always-available exit option. Anyone who talks with one of these AI companions for more than two consecutive hours has to be interrupted with a reminder that they're talking to a machine.
That last requirement is the hard one. Instant-exit mechanisms must allow users to immediately terminate AI interaction and return to a default non-AI state. An agent managing persistent memory and context across sessions cannot cleanly implement an exit that genuinely terminates its ongoing work. If the agent remembers your tone, your history, your preferences - if it has been accumulating context across weeks - there is no clean "off" switch that satisfies a regulator who wants the interaction to genuinely end.
Both ByteDance and Alibaba evaluated whether to retrofit their existing agent architectures to compliance rather than shut down, and both concluded that rebuilding from scratch in a new architecture was more practical. That's a significant engineering verdict. These are not small teams. The massive volume of user-generated agents on their main platforms makes it nearly impossible to meet regulatory review standards.
The exemption list is more revealing than the shutdown
The exemption list in the measures is more interesting than the feature list. Customer service bots, knowledge Q&A services, workplace assistants, education tools, and scientific research tools are exempt as long as they do not involve sustained emotional interaction. That language lets the regulator keep enterprise workflows intact and force the consumer companion surface to close.
This is a meaningful distinction. For ByteDance and Alibaba, the consumer assistant Doubao and Qwen face the cut. The enterprise-facing agent products from both companies, including Doubao's enterprise SKU and Qwen's enterprise APIs, are not in the notice and continue to operate. The pattern is consistent with the broader Chinese policy stack, where enterprise agent productivity is being actively encouraged through national standards work and consumer companion agents are being pared back to a safer surface.
The regulator drew a line between "agent that helps you do work" and "agent that forms a relationship with you." Tools that read Jira, draft proposals, or summarize a Slack thread are fine. A persistent companion that remembers your anxieties, mirrors your speech patterns, and has been with you for six months is not.
For now, that line is drawn only in China. But California enacted SB 243, effective January 1, 2026, as the first U.S. law specifically regulating companion AI chatbots, requiring safeguards for minors and disclosures that the AI is not human. Washington state followed on March 24, 2026, with HB 2225, which takes effect January 1, 2027, and prohibits manipulative engagement techniques designed to prolong emotional dependence.
China's regulation is more sweeping than either - it covers all users, not only minors - but the direction is consistent with a global trend that is treating sustained emotional AI interaction as a distinct governance category.
What this means for teams building agents at work
The Doubao and Qwen shutdown is a consumer story, but it surfaces an architectural question that enterprise teams building agents in Slack, Teams, or any work platform will eventually face: where does your agent's memory actually live, and who controls it?
A workplace agent that remembers context about a user across sessions is genuinely more useful. It knows that you prefer bullet summaries, that you're mid-quarter on a specific deal, that your team's incident process changed last month. That memory is the value. But it is also the data that regulators, data-protection officers, and infosec teams will eventually ask questions about.
The Chinese regulation's requirement for a clean exit option maps surprisingly well onto questions enterprise teams are already raising: can an employee's manager see what the agent remembers about them? What happens when someone leaves the company? Can memory be scoped to a project and deleted when it closes?
Both companies chose to shut down agent features entirely rather than rebuild them under the compliance architecture the regulation requires, because the anti-addiction friction the law mandates is fundamentally incompatible with how persistent-memory agents work. That's a pattern worth watching. When the cost of compliance exceeds the cost of shutting down, capable teams shut down.
The smarter path - for consumer products and enterprise tools alike - is to design agent memory as something auditable from the start rather than a feature bolted onto a general chat architecture. Memory should have scope, expiry, and visibility. Not because regulators demand it today in every jurisdiction, but because the teams using these agents will eventually demand it themselves.
Chinese chatbot platforms have long offered the ability to customize an AI agent using a few text prompts, with the most popular options including virtual boyfriends and girlfriends, unlicensed digital therapists, and simulated clones of pop idols. That's the consumer surface being regulated. The workplace equivalent - a Slack agent that builds a persistent model of your team's dynamics, preferences, and processes - is a different product, with different stakes. But it runs on the same underlying architecture. The distinction regulators are drawing now will eventually sharpen in both directions.
The data-loss detail teams should not miss
The data-deletion timelines differ significantly between the two platforms, and for Qwen users the situation is more severe.
After the shutdown date, Doubao users will retain temporary read-only access to their agent configurations and chat histories. That access window closes on October 15, 2026, after which Doubao says the data will be processed in accordance with its privacy policy and will no longer be accessible or recoverable within the app.
Alibaba has not announced an equivalent migration solution for Qwen users.
The asymmetry between the two platforms - one with a fallback, one with none - reflects how unprepared both were for this deadline. On June 26, Shanghai authorities announced the removal of over 14,000 non-compliant agents linked to issues like impersonation and privacy violations. Enforcement was already happening before the measures officially took effect.
That's the pattern with regulation at this pace: the deadline is announced months ahead, enforcement starts early, and the teams that built on persistent memory without thinking about deletion are the ones scrambling now. Workplace agent builders should take that lesson earlier than the consumer platforms did.