Executive Summary

On July 15, 2026, China's new rules take effect. ByteDance's Doubao and Alibaba's Qwen will shut down the feature that let users build their own AI agents, and they will delete the conversation histories those agents accumulated. Qwen has stated plainly that deletion is immediate and permanent, with no migration path. This article reads the event not as content regulation but as a question of data ownership.

What disappears is not the service but the memory. An agent's worth lives in the relational data built up through long exchanges with a user, and the rules turned exactly that data into something to be erased. The regulation did not order deletion directly. Platforms decided that shutting down was cheaper than compliance, and memory is being wiped as a byproduct of that decision.

Debates over AI data ownership usually stop at training data. Yet the core data of the agent era is the memory that accumulates during operation. An agent that was never designed for a data lifecycle and portability from the start can vanish wholesale at the stroke of a single rule.

The mark the rules leave on user data compresses into four numbers: when they take effect, how open the migration path is, how much has already been erased, and where the boundary of the rights left to users actually falls.

7.15

Rules take effect

Doubao and Qwen shut down agent features

0

Qwen migration paths

Immediate permanent deletion, no migration

14k+

Agents pre-emptively deleted

Shanghai authorities, a month before effect

Copy & delete ✓

Portability ✗

The rights Article 16 gives and withholds

1

What Disappears on July 15

Doubao closes its custom-agent feature on July 15. Until now, the conversations a user held with an agent they built stayed as read-only until October 15, after which they are processed under the company's privacy policy and permanently deleted. Qwen moves faster. It disables the feature on July 10, shuts it down entirely on the 15th, and deletes the data immediately and permanently. Neither company offers a way to export chat histories in structured form. The most a user can do is capture the screen or copy the text.

This is not just about two companies. Tencent's Yuanbao already retired similar features in June, and a month before the rules took effect, Shanghai authorities pre-emptively deleted more than 14,000 non-compliant AI agents. In line with a single line drawn by regulation, an entire industry moved in the same direction.

Data Deletion Timeline Qwen July 10 feature disabled July 15 shut down immediate deletion · no migration Doubao July 15 feature closed until Oct 15 read-only Permanently deleted no export path
▲ Doubao and Qwen agent feature shutdown and data deletion schedule | Pebblous original diagram

The complaint Doubao users left behind reduces to one sentence: "There is no easy way to export my conversation history." What they lose is not a handful of chats but an entire persona they raised by hand, setting its character, tone, and expertise. And the memory that persona built up while coming to understand the user disappears along with it.

2

What the Rules Targeted

The formal name is the Interim Measures for the Administration of AI Anthropomorphic Interactive Services. Five agencies, including the Cyberspace Administration of China (CAC), promulgated them on April 10 and they take effect on July 15. The rules apply to services that mimic a human's personality, way of thinking, and manner of conversation to form an ongoing emotional relationship with the user. Customer-support bots, knowledge Q&A, and work assistants fall outside the scope as long as they do not form that kind of relationship.

The key point is that the rules never directly ordered anyone to "delete memory." What they mandate is anti-addiction systems, usage-time reminders, and off-ramps that let a user leave whenever they want. But a design that remembers a user over time and holds them close emotionally, and a design that creates friction so a user can leave easily, are hard to satisfy at once inside a single service. Platforms weighed the cost of rebuilding from scratch against the cost of closing the feature, and chose to shut down. Memory is erased as a consequence of that choice.

Article 16 of the rules gives users the right to copy, delete, and manage their interaction data. It also bans using sensitive interaction data for model training without separate consent. But the right to portability is not on that list. Users have the right to erase their data, yet no explicit right to move it elsewhere and keep using it.

Article 16: Rights Granted vs. Missing ✓ Copy users may request a copy ✓ Delete users may request deletion ✗ Portability not specified — absent
▲ User rights structure under China's AI anthropomorphic regulation Article 16 — portability not included | Pebblous original diagram

Here lies the paradox of the event. Users are given the right to copy and delete, yet in the very situation where that data vanishes wholesale, no destination is provided to move it to. A bundle of copied text cannot make another agent inherit it and remember the user. The form of the right exists, but the means to preserve the relationship is empty.

3

The Value Lives in Accumulated Memory

Why does memory matter so much? As of 2026, agent memory is no longer a side feature but a first-class architectural component. Dedicated benchmarks have appeared, a distinct research field has emerged, and a tooling ecosystem has grown up around it. Memory is composed by weaving together vector databases, graph relationships, and key-value stores, and at the start of each session it retrieves relevant memories and injects them into context.

The crux is that this memory is not a plain conversation log. It is structured understanding that compounds over time. An agent attaches a scope to each memory: some used only within a session, some shared across an organization, and some bound to a specific user and persisting beyond a single session. Among these, the memory bound to a user is precisely the relational data. It is why an agent feels as though it "knows" me, and why I no longer have to explain everything from the start each time.

This trend has already descended to the infrastructure layer. AWS Bedrock, Google Vertex AI, and Cloudflare have each released agent memory as a managed service. The work of extracting information from conversations, storing it, and reviving it when needed is shifting from a feature you build yourself to a service the cloud sells. Memory has become that valuable an asset.

Agent Memory Scopes User-bound Memory persists across sessions · relational data · target of deletion under China's rules Organization-shared Memory shared knowledge across team or org · available across sessions Session Memory valid within a single conversation only
▲ Agent memory scopes — user-bound memory (outermost layer) is what China's rules targeted for deletion | Pebblous original diagram

If so, what the rules turned into something to be deleted is not the shell of the agent but its core. The relational data that user and agent built together — the very layer that made the agent theirs alone — is what disappears. A service can be rebuilt, but that relationship has to be accumulated all over again.

4

An Agent Without a Data Lifecycle Barely Exists

When we talk about AI data ownership, we mostly picture training data. What text was the model trained on, and how do copyright and consent for that data stand? But once an agent actually starts working, the ownership question shifts from before training to after deployment. Whose is the memory that piles up every day? When a service closes its doors, where does that memory go?

Gaps show even in existing regulatory frameworks. GDPR Article 20 specifies a right to data portability, but there is no standard for what format an agent's memory graph should be moved in. The EU AI Act demands transparency and logging but does not address auditing a user's memory. China's new rules grant copy and delete rights while leaving portability blank. No framework answers the question of whether a user can carry out the memory accumulated during operation.

Those who have long worked in data governance already know the direction of the answer. Set retention and deletion policies before memory starts accumulating; ship, from day one, tools that let users inspect, correct, and delete their own memory; make memory exportable in a general-purpose format such as JSON; and store the original text alongside the vectors so that migration remains possible. The problem is that only a minority of platforms have actually adopted these practices. Neither Doubao nor Qwen managed to.

Qwen's "no migration path," then, is more than a notice line. It is a declaration that the user's relational data evaporates completely. An agent that never built a data lifecycle and portability into its design runs perfectly well day to day, then loses the entire value it accumulated at a single external shock — a regulation, or a business shutdown. From the standpoint of memory, such an agent is not much different from one that never existed at all.

The agent that survives longest under a regulatory environment is not the one with the best performance, but the one that carries a data lifecycle in its body from the start. How memory is accumulated, how it is handed back to the user, and how it is moved elsewhere must all be in the blueprint. This is, before it is a matter of compliance, a matter of protecting the value the agent created.

Editor's note. Pebblous has long worked on the process by which data becomes ready to be used by AI (AI-Ready Data). This event shows that the scope of that readiness is widening from training data to the memory that accumulates during operation. Only when agent memory, too, is managed for quality, lifecycle, and portability can it be called AI-Ready.

Frequently Asked Questions

What are China's rules on AI anthropomorphic interactive services?

They are rules promulgated by five agencies, including the Cyberspace Administration of China (CAC), on April 10, 2026, taking effect on July 15. They target AI services that mimic a human's personality, way of thinking, and manner of conversation to form an ongoing emotional relationship with the user, and they mandate anti-addiction systems and off-ramps.

What happens to my chat history when Doubao's agent feature shuts down?

Doubao shuts down the feature on July 15, keeps existing conversations read-only until October 15, and then permanently deletes them under the company's privacy policy. There is no structured export; you can only capture the screen or copy the text.

Why can't Qwen agent data be migrated?

Alibaba's Qwen shuts down fully on July 15 and deletes data immediately and permanently, officially stating there is no migration plan. Because the rules grant copy and delete rights but do not specify a portability right, there is no institutional path at all for users to move their memory to another service.

Why is an AI agent's memory worth more than a simple log?

Agent memory is not a plain record of conversation but structured understanding that compounds over time. Memory bound to a user is the relational data that makes an agent theirs alone, and personalization, trust, and utility flow from it. As of 2026, agent memory has become a first-class asset that the cloud sells as a managed service.

Who owns the data in an agent's memory?

For now, it is effectively locked to the platform. Users can demand a copy or deletion, but with no standard to move memory elsewhere in its original context, when the platform disappears the memory disappears with it. This article argues that to genuinely return ownership to users, portability must be built into the design.

Could rules like these spread to Korea or other countries?

The form of regulation will differ by country, but the emotional relationships and data issues around AI companions and agents are already a global concern. California runs a deletion-request regime, and the EU is strengthening the right to data portability. Even where the direction of regulation differs, the question of "how to handle memory accumulated during operation" remains common to all.

How should I account for the data lifecycle when designing an agent?

It is recommended to set retention and deletion policies before memory accumulates, to provide tools for users to view, edit, and delete their own memory from day one, to make memory exportable in a general-purpose format such as JSON, and to store original text alongside vectors so migration remains possible. Agents built with these practices withstand regulatory and business change.

How do China's rules and Europe's GDPR differ on AI agent data rights?

GDPR Article 20 specifies a right to data portability, but there is no concrete standard for moving an agent's memory graph. China's new rules grant copy and delete rights while leaving portability blank. Both frameworks mention the principle of deletion or movement, but neither yet has a practical standard for how to migrate the agent memory accumulated during operation.

R

References

News & Press

Legal & Policy Analysis

Technology & Industry

Academic