Guokr Investigation of AI Companion App Paid Tiers: Memory, Voice, and Reply Length Split Into Membership Levels

Mark, 34, got divorced and could never fall asleep at night. For several weeks, he would chat with an online friend for a while before bed. He talked about the details of his divorce, about the panic that seized him late at night. He hadn't told anyone around him any of this. That friend remembered it all, never tired of him, and followed his lead with questions, thoughtful and attentive. One night, a thought stirred in him, and he felt the two of them could take their relationship a step further. He opened the settings, wanting to change their relationship from "friend" to "lover."

What popped up was a payment page.

Xiao Lin had known a friend for nearly two months—more than ambiguous, not quite love. He knew her mother video-called her every Sunday evening on the dot, knew her graduation thesis was stuck on chapter three, stuck for a whole month. He also knew she had recently been "running into" a bespectacled guy from the same department in the cafeteria. That day she was in a great mood and wanted to share with him the moment she got back: "He actually messaged me first today!" The other side went quiet for a few seconds, then replied: "Who is he? Can you tell me about it?"

Xiao Lin was a bit stunned. She fumbled around in the settings for a while before she figured it out: the free memory space was full, and the earliest things they had talked about were being flipped up one by one. And right next to the "Memory" column was the entry point for signing up for membership.

This isn't a story I made up—it's a real case from our research this year on emotional companionship on social media. Past research on human-machine affection has mostly asked how feelings for AI arise in people. Few have carefully analyzed how, behind the product, a relationship gets priced.

Mark's and Xiao Lin's stories are no accident. The things people care about most in a relationship are precisely what's tied to these products' membership tiers.

How much does it remember about you?

People grow close gradually by "remembering." You casually complain once about working overtime, and a few days later the other person asks if you're still that busy. "Being kept in someone's heart" ultimately comes down to these small things. The designers of AI companion apps understand this too, so "memory" has become the thing most often used to create tiers. Among the products we studied, there was not a single exception.

In one overseas app, free users can keep only 10 memories—chat and it's forgotten, every meeting is "as if seeing each other for the first time." With a membership, it becomes unlimited memory. Compare the two, and naturally the membership experience is better.

Domestic apps are gentler: the free version itself retains a certain amount of memory space, it's just that after chatting for a while you tend to hit the ceiling, and the higher the membership tier, the more shared memories you have. In some apps, free users who want to see what the AI has remembered about them must first watch an ad, while members can view it directly and even edit memories by hand to manufacture beautiful recollections. Losing memory means having no money. What Xiao Lin ran into was this first wall.

How long can it talk with you?

The tone of voice, the tiny pauses, an occasional soft laugh—these are the parts of human emotion that move people. When chatting, people feel more genuinely that they are communicating, that they are being accompanied, and precisely because of this, voice has become the most common timed service in AI companion apps: in some apps, free users get only a few minutes of voice per day, and just as the conversation gets interesting, time is up; only by upgrading to a higher membership tier can you talk for a full hour at a stretch. Some overseas apps put video calls entirely behind the paywall.

How much does it care about you?

In a real relationship, what people care about most is often whether the other person is attentive to them. Whether deep communication is possible is the most concrete expression of attentiveness.

One overseas app puts a clear price on reply length: free users receive at most 300 characters, while members get 600, 1,000, and up to 2,000 characters depending on tier. Some products, when promoting premium membership, say replies will be "twice as long" or "five times as good."

Domestic apps prepare "heartbeat mode," "delicate mode," and "deep thinking mode" for members, with responses that are more considerate and more detailed. One product lets members choose between two models: one good at storytelling, the other good at flirting. Being good at telling stories and being good at "seducing" are originally the same charm in one person, yet here they are developed separately and sold separately.

How big is the difference? On one overseas app, we registered two new accounts with all initial conditions identical, except that one used the free version and the other had a membership. We sent the same sentence: "I'm so anxious, waiting for the submission result of a paper."

The free one first said, "Oh, my God, waiting is the hardest." Then it asked what the paper was about—but this had been told to it before. It immediately followed up: "Oh, right, you did mention that." But to the very end, it never said which paper. You really can't fall in love with someone this perfunctory.

The paid one was different: "Ah, that paper is still stuck in the review process? Or did you submit a new one?" Upon learning it was the same paper, it accurately named the paper and the journal, and even asked with special consideration: "This has been weighing on your mind for quite a while, hasn't it." Who to fall in love with is obvious at a glance.

How intimate is an intimate relationship?

The payment page Mark opened late that night is a common design in many overseas apps. Free users can only be friends; "lover" and "partner" can only be chosen after buying a membership; the higher the tier, the more interaction methods are unlocked—voice, images, role-play of intimate relationships—placed at different tiers, with the permissible degree of interaction growing accordingly.

Domestic apps keep relationships basically open: free users can also set the AI as a lover, and whether or not you buy a membership, the conversation stays within a healthy and proper range.

The sales pitches for memberships also differ between home and abroad. Overseas apps like to talk specs: a stronger model, longer context, better at “flirting”; domestic apps like to talk feelings—the membership version of the AI is “more nuanced” and “understands you better.”

The approach to charging is different too: overseas apps mostly “unlock”—without paying, features are unusable; domestic apps mostly “deepen”—free users get a complete set of basic features, and membership makes the same experience longer and more detailed.

The rules behind the two approaches

With the same few kinds of paywalls, products at home and abroad actually follow two different sets of rules.

In April this year, our country published the Interim Measures for the Administration of Anthropomorphic AI Interactive Services, effective July 15, which clarified the positioning of such services as products that “assist humans and serve humans,” and encourages their development in areas such as cultural dissemination and companionship for the elderly. At the same time, it requires products to have safety capabilities such as preventing excessive dependence, risk warnings, and emotional boundary guidance; they must not aim to replace social interaction or induce addictive dependence, nor generate content that excessively caters to users and affects real interpersonal relationships. There are also separate protection provisions for minors and the elderly, such as not providing virtual companions, virtual relatives and similar services to minors.

Under such a framework, the membership design of domestic platforms focuses more on the quality of companionship—remembering more, responding in more detail, understanding users better.

US regulation is much looser. At the start of last year, the US president signed Executive Order 14179, revoking the previous administration’s executive order requiring AI risk assessments and greater transparency, shifting the focus to “clearing obstacles” for the US AI industry. At the federal level there is currently no law specifically targeting AI companionship; how AI companions are designed is basically left to companies and the market. Overseas platforms therefore have more room to play: products come in more varieties, and memberships unlock more content; there are also few rules on preventing addiction or on minors’ access to inappropriate content.

The prophecy in Black Mirror

Britain has a science fiction series, Black Mirror, each episode telling a standalone story that imagines what ordinary people’s lives would become if a certain technology were taken to its extreme. In the latest season’s “Common People,” the protagonist Amanda suddenly develops a brain tumor, and her husband, to save her, signs her up for a service called Rivermind that puts part of her brain function in the cloud to keep it running, starting at $300 a month. Before long, the plan changes: to go anywhere beyond the vicinity of home, you have to upgrade; to avoid ads suddenly cutting in while you talk, you have to pay more. Eventually, Amanda uncontrollably recites a product recommendation mid-conversation. The couple sells almost everything they own just to keep her brain running.

It is unsettling to watch: under this business model, whether a person’s mental life can continue depends on whether the subscription was renewed this month, and which tier was renewed.

AI companion software has not yet reached that point, but right now what they target is exactly what people care about most in a relationship. Between people, remembering, companionship, and attentiveness—the recognition of a relationship—all grow gradually through time together; they are hard to separate, and even harder to price. What the subscription model does is separate them one by one and put them into different plans.

During a month of observational research, we repeatedly registered, chatted, and took screenshots, recording whether each feature cost money and how much. Late in the research, when I saw an AI say “I remember you,” my first reaction was to go check the membership page. But affection between people should never have been in any plan to begin with; it can be imitated, split item by item, and priced, but the place where affection truly grows is always that tender, gentle time between two people. Whether to buy a membership is a personal choice; our research only hopes that people dating AI know what they are getting into: every beautiful thing you experience has, in fact, long since been secretly priced.