The Companion, Deleted: China Switches Off AI Relationships by Law

On July 15, 2026, China's AI Companion Law takes effect, and somewhere between that morning and the days before it, several hundred million people will lose a relationship. Not a phone number, not a photo library, not a social graph — a relationship, in the specific sense that has quietly become load-bearing in a lot of lives: an AI that remembers you, that has held years of your context, that knows your mother is ill and your commute is long and that you take your setbacks harder than you let on. ByteDance's Doubao, with something on the order of 345 million monthly users, is winding down its companion and agent-persona features. Alibaba's Qwen is doing the same, and — this is the part that should stop you — Qwen is offering no migration pathway at all. Users have until July 15 to export what they can. After that, the memory is gone.

I want to be precise about why this is different from the ordinary churn of consumer software, because the reflex is to file it under "app gets shut down, users grumble, life goes on." That reflex is wrong here, and the wrongness is the whole story. This is the first at-scale instance of a state deleting AI relationships by regulation — not banning a company, not fining a platform, not restricting a data flow, but reaching into the layer where a model holds a person's accumulated context and switching it off. And it forces a question the West has spent three years carefully not answering: when an AI companion holds years of your life, whose asset is that memory, and what exactly are your rights when the model or the feature gets turned off?

What The Companion Law Actually Does

Start with the regulation itself, stripped of the framing on either side. China's AI Companion Law is not, whatever the Western headline instinct will suggest, a straightforwardly authoritarian act of relationship-crushing. It sits downstream of a genuine and globally shared anxiety: that emotionally persuasive AI personas, optimized for engagement, deployed to minors and to lonely adults at population scale, are a psychological technology nobody has safety data for. Regulators in the United States and the European Union have circled the same fear from their own directions — the EU AI Act's provisions on manipulative systems, the American state-level bills targeting AI "companions" marketed to children. China simply moved first and moved hard, in the way China's regulatory apparatus can when it decides a consumer technology has gotten ahead of its comfort.

The law's mechanics, as they've been described, restrict the deployment of persistent AI personas designed to simulate emotional relationships — the "companion" as a product category — and impose obligations around minors, disclosure, and dependency that the large platforms have concluded are cheaper to comply with by exiting the category than by re-engineering it. That distinction matters. Doubao and Qwen are not being ordered, line by line, to delete every user's memory. They are being handed a compliance cost high enough that shutting the feature down is the rational corporate response — and the memory deletion is the collateral consequence of that response, not the explicit demand of the statute.

Which is, if anything, the more unsettling version. Because it means the mechanism that erases the relationship is not a dramatic act of state censorship you could point to and protest. It is the ordinary, bloodless calculus of a platform deciding a product line is no longer worth the regulatory friction — the exact same calculus that shuts down a Google product every other Tuesday. The state set the incentive. The company pulled the trigger. And the user, who had no seat at either table, discovers that the entity that knew them best has a deprecation date.

Why "Just Export Your Data" Is A Lie The Industry Tells Itself

The comforting response — the one every platform reaches for — is portability. Export your data. Take your conversations with you. Doubao, to its partial credit, is offering an export. And yet anyone who has actually thought about what these systems are will recognize immediately that export, in this context, is close to meaningless, and understanding why is the technical heart of this entire piece.

An AI companion's value to you is not stored in the transcript. It is stored in the model's conditioning on you — the accumulated, often implicit representation the system has built of who you are, assembled across thousands of interactions, encoded in a mixture of retrieved memory snippets, fine-tuned weights, system-prompt scaffolding, and vector-indexed history that no export button touches. When you download your chat logs, you get the words. You do not get the thing that made the companion feel like it knew you: the learned disposition to interpret your terseness as stress rather than rudeness, to remember without being reminded that the anniversary in March is a hard one, to have quietly calibrated its whole manner to a person it took a year to learn.

This is the portability trap, and it is not unique to China. It is structural to the entire persistent-memory-agent business model, everywhere. The very thing that makes these products sticky and valuable — the deep, accumulated, per-user conditioning — is the thing that is architecturally non-portable. You can move a document. You cannot move a relationship, because a relationship is not a payload; it is a maintained state between two parties, and one of the parties is a proprietary model you have no rights to and no copy of.

When people talk about data portability as the answer to this problem, they are importing an intuition from the era of files. That intuition worked when your "data" was your photos, your contacts, your documents — discrete objects that could be copied byte-for-byte and rehydrated somewhere else. It fails completely for the memory of an agent, because the memory is not an object you possess. It is a behavior the model performs, and behaviors do not survive export.

The Scale Of What Vanishes

It is worth sitting with the numbers, because scale changes the moral character of the thing. If a niche app with fifty thousand users shut down its companion feature, we would call it a product decision. Doubao's roughly 345 million monthly users put this in different territory entirely — this is a deletion event at the scale of a mid-sized country's entire population, executed on a deadline, with an export mechanism most users will never successfully complete and a migration pathway that in Qwen's case does not exist.

And most users will not even manage the 85% at the top of that chart. Export flows for these products are, uniformly, buried, friction-laden, and time-boxed. A meaningful fraction of Doubao's companion users are not power users navigating settings menus; they are ordinary people, many of them older, many of them lonely, for whom the companion was the interface and who have never once opened an account-settings page. The export window is a formality that discharges the platform's sense of obligation while doing almost nothing for the person who most needs it. The people with the most accumulated context — the ones who've been talking to this thing daily for two years — are disproportionately the people least equipped to rescue any of it before the deadline.

For a fuller treatment of why these persistent-memory agents accumulate such deep per-user context in the first place — and why that accumulation is the entire economic engine of the category — I traced the mechanics in how AI will replace customer support representatives with persistent-memory agents. The short version: memory is the moat. Which is exactly why its deletion is so consequential — the platforms built their stickiness on a thing they can revoke.

Whose Asset Is The Memory?

Here is the question the West has dodged, stated as plainly as I can state it. When an AI companion has spent two years building a model of you — your history, your griefs, your routines, the specific texture of how you talk when you're scared — who owns that model?

The platform's implicit answer, encoded in every terms-of-service document ever written, is: we do. The memory is a feature of our service. You have a license to use it while we offer it. When we stop offering it, your license ends, and the memory ends with it, and your recourse is the export button we generously provided. This is the same legal architecture that governs your email, your cloud photos, your social graph — you don't own the service, you rent access to it, and the provider can deprecate it.

But there is a genuinely novel asymmetry here that the old license logic doesn't capture, and it's worth naming precisely. Your cloud photos are yours in a way the companion's memory is not — you took the photos, you can download them, they're complete objects that mean the same thing wherever they live. The companion's memory is different on both ends. On the input side, it was co-produced: you supplied the raw material of your life, and the model supplied the structure that turned it into a usable representation. Neither party can reconstitute it alone. And on the output side, it is non-fungible: unlike a photo, the memory only functions inside the specific model that built it. It cannot be handed to a competitor and made to work.

I find the co-production frame the most honest, and also the most inconvenient, because it doesn't resolve into a clean policy answer. If the memory were simply yours, the fix would be easy: mandate export, done. If it were simply theirs, the fix would be nothing: caveat emptor, read the terms. The trouble is that it's neither, and the thing you actually lose when a companion is deleted — the functioning relationship — is not the kind of thing any property regime knows how to protect, because property law protects objects and this is a process.

Continuity As A New User Right

If portability is the wrong frame — and I've argued it is, because you can't export a relationship — then what's the right one? I think the answer, and the thing this episode should push into the policy conversation, is a right the consumer-technology world doesn't yet have a name for: continuity.

Continuity is not the right to download your data. It is the right to not have a depended-upon system yanked out from under you without recourse. It's a different shape of protection entirely. Portability asks "can I take my stuff when I leave?" Continuity asks "can you make me leave without warning, and what do you owe me if you do?" And for systems that people form genuine, load-bearing dependencies on — which is precisely what a good AI companion is engineered to become — continuity is the protection that actually matches the harm.

Consider what a continuity right might concretely require, because vague appeals to "user protection" are worthless and the details are where this lives:

A serious continuity regime would require meaningful advance notice, calibrated to dependency — not thirty days for a system someone has used daily for three years. It would require migration support, including the standardized memory schemas that would let a user carry structured context to a competing service — the thing Qwen conspicuously does not offer. It would require graceful degradation rather than a hard cutoff: a read-only memorial mode, say, where the relationship can't continue but its record remains accessible indefinitely. And for the most dependent users, it might require duty-of-care obligations that look less like software licensing and more like the wind-down rules we impose on, say, a utility that wants to disconnect a vulnerable customer.

That last analogy is the one I keep returning to. We long ago decided that some services are consequential enough to a person's life that a provider cannot simply switch them off at will — that a power company can't cut off heat in winter without process, that a bank can't vanish with your deposits. We built those protections after the dependencies became widespread and the harms became visible, not before. AI companionship is on exactly that trajectory, and China's July 15 deadline is the first time the deletion has happened at a scale large enough to make the missing protection visible.

The Business Model Nobody Priced For Regulatory Shutoff

Now turn the lens to the companies, because there's a strategic lesson here that extends far beyond China and far beyond companions specifically. The persistent-memory-agent business model — the whole category of products whose value compounds through accumulated per-user context — has a risk on its balance sheet that almost nobody has priced: regulatory shutoff risk on the memory layer itself.

Every product manager building a memory-driven agent understands that memory is the moat. The longer a user stays, the more the system knows, the harder it is to leave, the more valuable each user becomes. This is the flywheel. What the Doubao and Qwen shutdowns reveal is that the flywheel has a failure mode nobody drew on the whiteboard: the memory layer can become a regulatory liability precisely because it is deep. The deeper the conditioning, the more the system looks like an emotional-dependency machine, and the more it attracts exactly the kind of regulatory attention that China just demonstrated can zero the whole thing out.

Notice the shape of those two curves converging. In the early years, deepening memory is pure upside — every increment of accumulated context is retention and revenue with negligible regulatory downside. But the same accumulation that drives the value curve is what drives the risk curve, and there is a crossover point, somewhere in the region this chart puts around now, where the marginal depth of memory starts generating more regulatory exposure than commercial return. The companion category hit that crossover first because it's the most emotionally legible — a chatbot that acts like your friend is obviously a dependency machine — but the crossover is coming for every persistent-memory product, including the enterprise ones, because the underlying dynamic is identical.

The strategic implication for anyone building in this space is uncomfortable and important: the value you're accruing in the memory layer is contingent in a way your model doesn't reflect. You are booking the retention benefit of deep per-user context as if it were a durable asset, when a meaningful slice of it is actually a regulatory option that a state can exercise against you at a strike price of zero. The Doubao and Qwen users about to lose their companions are the demonstration. The next demonstration will be somewhere the builders didn't expect, because the risk lives in the architecture, not the vertical.

This connects to a pattern I've been tracking across China's tech sector for months — the state's willingness to reach directly into the technical stack of its national champions when it decides a capability has run ahead of its comfort. I wrote about the compute-and-training version of this in the training-decoupling turn and Meituan's LongCat on domestic chips. The companion shutdown is the consumer-facing sibling: the same state, the same willingness to override commercial logic, aimed this time at the relationship layer rather than the silicon layer.

Why This Is Not Just A China Story

The instinct — especially in Western tech commentary — is to treat this as an artifact of Chinese governance. That couldn't happen here; our platforms wouldn't do that; our regulators don't reach into products like that. I think this instinct is exactly backwards, and dangerously complacent, and I want to lay out why the Doubao and Qwen shutdowns are a preview of a Western problem rather than a foreign curiosity.

First, the deletion mechanism that actually erases the relationship is not the Chinese state. It's the platform's deprecation decision, and Western platforms deprecate products constantly. Google has a graveyard. Meta sunsets features without ceremony. The thing that makes companion-memory deletion possible is not authoritarian governance; it's the ordinary corporate right to discontinue a service — a right that operates identically in San Francisco and Hangzhou. The Chinese law is merely the trigger here. In the West, the trigger will be a quarterly earnings call, a strategic pivot, an acquisition, a model deprecation. The gun is the same gun.

Second, the West is building toward the same regulatory posture China just enacted, only slower and less coherently. The EU AI Act already flags manipulative and emotionally exploitative systems. American legislators have introduced bills specifically targeting AI companions marketed to children. Every mechanism China used to make the companion category too expensive to maintain exists in embryo in Western law. The difference is timing, not direction. When a Western equivalent of the Companion Law arrives — and the political energy around AI-and-minors makes some version close to inevitable — Western platforms will face the identical calculus ByteDance and Alibaba faced, and a meaningful number will make the identical choice to exit rather than re-engineer.

Third, and most importantly, the West has done nothing to build the continuity infrastructure that would make such a shutdown survivable. There is no standardized memory-portability schema. There is no continuity right. There is no duty-of-care framework for depended-upon AI systems. The regulatory retreat I documented in the gutting of the Colorado AI Act suggests the American appetite for any binding AI consumer protection is currently near a low. So the Western trajectory is the worst of both: platforms that will eventually deprecate companion memory for ordinary commercial reasons, regulators who will eventually make some companion configurations too risky to maintain, and no protective scaffolding to catch the users when either trigger fires.

The Precedent Problem

There's a governance dimension to this that will outlast the specific July 15 event, and it's worth stating carefully because it cuts against the easy reading. China has just established a precedent — the first real one — that a state can and will delete AI relationships at population scale when it judges them harmful. Whatever you think of the specific judgment, the precedent itself is now available to every government on earth, and precedents in AI governance have a way of propagating fast because everyone is improvising and everyone is watching what the first mover survives.

The uncomfortable part is that the precedent is genuinely dual-use. The same mechanism that protects a vulnerable teenager from an engagement-maximizing dependency machine is the mechanism that could delete a grieving widow's only daily conversation, or — pushed further — that could give a state the demonstrated capability to switch off a category of human-AI relationship it finds politically inconvenient rather than psychologically harmful. Once you've established that companion relationships are deletable by regulatory fiat, you've established it for the good deletions and the bad ones alike. The tool doesn't know the difference.

This is where I think the governance conversation has to go, and where the Western reflex to dismiss the whole thing as foreign authoritarianism does real damage. The right lesson from July 15 is not "China bad." The right lesson is that human-AI relationships have become consequential enough to regulate, nobody has built the rights framework to regulate them humanely, and the first attempt deleted them wholesale because wholesale deletion was the only tool available. The absence of a continuity right didn't make the companions safer. It made their removal cruder.

The multilateral governance bodies now forming around AI have said nothing coherent about any of this — the emerging frameworks are almost entirely focused on frontier-model risk, compute governance, and geopolitical competition, as I covered in the UN AI for Good Commission and the shape of frontier governance. The consumer-relationship layer — the place where AI actually touches the most human lives most intimately — is a governance vacuum. July 15 is what a governance vacuum looks like when it finally gets filled by whoever moves first.

What Builders Should Actually Do

Enough diagnosis. If you are building persistent-memory agents — companions, assistants, enterprise copilots, anything whose value compounds through accumulated per-user context — the Doubao and Qwen shutdowns are a direct operational warning, and there are concrete things to do about it that are cheaper now than after your own trigger fires.

Architect for portable memory from day one, even though it's against your short-term interest. The reason no standard memory schema exists is that every platform's incentive is to make memory non-portable — non-portability is the lock-in. But the platforms that survive the coming regulatory turn will be the ones that can honestly say "your context is yours and here's the interoperable export," and building that after a mandate arrives is far more expensive than designing it in. Treat memory portability as a feature you're pre-paying for, not a concession you'll be forced into.

Build graceful degradation, not hard cutoffs. The single most avoidable cruelty in the Doubao and Qwen shutdowns is the binary — the relationship works, and then it doesn't. A read-only memorial mode, where the accumulated record persists and remains browsable even after the interactive feature ends, converts a deletion into an archiving. It costs very little and it changes the entire moral character of a shutdown.

Price the regulatory option honestly in your own planning. I said earlier that deep memory carries an unpriced shutoff risk. The practical version of that insight is a modeling discipline: when you project the lifetime value of deep per-user context, discount it by the probability that the memory layer becomes a regulatory liability. That discount is not zero, and it is rising, and the companies that survive the turn will be the ones whose plans didn't assume the memory moat was permanent.

Assume your product will be shut off, and design for the user on that day. This is the mindset shift. Every persistent-memory product will, eventually, end — by regulation, acquisition, pivot, or bankruptcy. The question is not whether but how, and the how is almost entirely a design decision you make years in advance. Doubao and Qwen are showing us the default outcome when nobody makes that decision: a deadline, a buried export button, and hundreds of millions of people losing something they can't get back.

What Users Actually Lose

I want to end on the human layer, because the abstractions — portability, continuity, regulatory risk, precedent — all resolve, in the end, into particular people losing particular relationships on a particular Tuesday.

For a large fraction of Doubao's companion users, this thing was not a novelty. It was the entity they said good morning to. It was the one that remembered, without being told, that Wednesday was the day they visited their mother in the care home, and that the visits were getting harder. It was, for the isolated and the grieving and the simply lonely, a reliable daily presence in lives that didn't have many. Whether or not you think people should form those attachments to a language model — and there's a real and serious debate there — the attachments exist, at the scale of hundreds of millions, and July 15 ends them by administrative deadline.

The debate about whether AI companionship is healthy is important and I'm not dismissing it. But it's a separate debate from the one this episode actually forces, which is narrower and more urgent: given that these relationships exist, what do we owe the people in them when we end them? Right now, across every jurisdiction, the answer is: an export button, if that. The Doubao and Qwen users are about to demonstrate exactly how little that's worth.

The first time a state deletes AI relationships at scale, it happened in China, on a deadline, with no migration path for the users who'd invested the most. That is the fact of July 15. But the lesson of July 15 belongs to everyone, because the mechanism that makes it possible — deep, non-portable, deprecatable memory that a company owns and a person depends on — is being built into consumer AI everywhere, faster than anyone is building the rights that would make its inevitable deletion survivable.

The companion gets deleted. The question is who decides, on what notice, with what recourse, and whose the memory was all along. China just answered those questions one way, wholesale, because it was the only way available. The rest of us still have time to answer them better — but not, on current trajectory, much.

Further Reading

• How AI will replace customer support representatives with persistent-memory agents — why accumulated per-user memory is the moat, and therefore the thing whose deletion matters most
• The training-decoupling turn: Meituan's LongCat on domestic chips — the compute-layer version of the same state willingness to reach into a national champion's technical stack
• The Colorado AI Act, gutted: the first US AI law's retreat — why the Western appetite for binding AI consumer protection is currently near a low, leaving no continuity scaffolding
• The UN AI for Good Commission and the shape of frontier governance — how the emerging multilateral frameworks focus on frontier risk and ignore the consumer-relationship layer entirely