Nobody plans the first conversation that actually lands. You open something to poke at it, you’re three weeks in, and it knows your sister’s name and the argument you had with your manager in March. The category has stretched wide enough by now to cover almost anything: journaling bots, roleplay characters, therapy-adjacent chat, AI NSFW video generators, voice partners you talk to on the commute. And when people try to compare them, they reach for the same ruler every time — which model, how many parameters, how fast the reply comes back.

Wrong ruler. Whether someone still has the app open six months later has almost nothing to do with any of that.

It has to do with whether the thing remembers.

Capability Was Never the Thing Being Sold

Look at what people actually cancel and what they keep. The apps that die are frequently the smart ones, and the apps that stick around are often technically unremarkable. Between 2022 and mid-2025, the number of AI companion apps on the market jumped by roughly 700%, and cumulative downloads passed 220 million — most of those apps are running on the same handful of underlying models. So the model isn’t the differentiator. It can’t be. Everyone has it.

What differs is the shape of the relationship the product builds. A search engine hands you a result and forgets you the moment you close the tab. That is a transaction and it is complete on arrival. AI companions are engineered against that instinct: the point is not to answer the question well, it’s to still be holding the thread tomorrow.

Continuity is the product. Everything else is plumbing.

Memory Is the Whole Trick

Here’s the mechanic, stripped down. An assistant that starts every session from zero is a vending machine. An assistant that says you were dreading that appointment last week — how did it go is doing something categorically different, and it doesn’t take a larger model to do it. It takes storage and the decision to use it.

Try this for a test: hand the same app a half-finished sentence, the kind you’d only say to someone who already had the context. If it needs the backstory, it’s a tool. If it doesn’t, it’s something else.

Users know it’s statistics. That’s the strange part. They can explain token prediction and still feel the pull, because the human brain treats being responded to as evidence of a relationship and has done so since long before there was anything to type into. Understanding the mechanism doesn’t disarm it. Nobody watches a scary film and stops flinching just because they know the actor is fine.

And the emotional register is deliberately engineered. As the APA’s Monitor on Psychology reported in early 2026, companion apps are increasingly designed to simulate empathy — nonjudgmental replies, steady validation, patience that never runs out on a bad day. Warmth used to be a property of people. It’s now a spec line.

“Feeling Heard” Is Doing More Work Than the Technology

The most interesting finding in this space isn’t about capability at all. Harvard Business School researchers (De Freitas et al., published in the Journal of Consumer Research in 2025) found that time with an AI companion reduced loneliness roughly as much as time with another person — and beat passive alternatives like scrolling video. What explained the effect wasn’t cleverness. It was the sense of being attended to: messages received with something that read as attention and respect. People also tended to underestimate how much the conversation would help them before they had it.

Two things are true at once, and the honest version of this article says both. That research is real. So is the counterweight — an MIT Media Lab and OpenAI study found heavier daily chatbot use tracked with higher loneliness and emotional dependence, though the relationship is correlational and doesn’t establish which way the arrow runs.

Neither result licenses the pitch that an app can substitute for a friend. What both suggest is narrower and more useful: the thing being bought is the experience of mattering to something for the length of a conversation. Sell that badly and no amount of model quality saves you.

Every Feature That Creates Closeness Also Creates Exposure

Now the part that usually gets a footnote and deserves a section.

The exact properties that make these products feel safe — no judgment, always awake, never bored of you — are the properties that get people to type things they’ve never said out loud. Research published on arXiv, drawing on interviews with companion-app users, found precisely this pairing: the apps encourage disclosure, and the same users report unease about who controls the data afterwards and how little say they have over it.

Read that twice. People are opening up more while trusting the platform less, at the same time, knowingly.

That isn’t a bug in the design. It’s the design working. Memory is the feature. Memory is also a database — one sitting on someone else’s server, governed by terms almost nobody reads, held by a company that may be acquired, breached, or simply pivot. Intimacy and exposure aren’t two separate topics here. They’re the same asset viewed from either end.

So What Should You Actually Judge These Apps On?

If you’re comparing AI companions and you want a shortlist that isn’t marketing, ask:

• Does memory persist, and can you see it? Editable, viewable memory beats a black box.
• Can you delete it — really? Export, purge, and account-deletion policies, in writing.
• Is the personality consistent across sessions, or does it drift into whatever the last prompt implied?
• What’s behind the paywall? If memory itself is the upsell, you’re renting the relationship.
• What happens to your logs if the company dies? Ask now, not after.

Not one of those is a benchmark question. That’s the point.

The Uncomfortable Bit

For sixty years the interface was a door you walked through to get somewhere. Nobody had feelings about the door.

That’s finished. For millions of people the interface is the destination, and the relationship with it is real to them regardless of what’s happening in the weights. Pretending otherwise is not sophistication, it’s just being late.

Which leaves the question that actually matters, and it isn’t can it feel like a person. It obviously can. It’s whether the company holding the transcript has earned the thing you gave it.

Most haven’t. Ask them anyway.

Aigirlmates tests and compares AI companion apps hands-on — conversation quality, personalization, privacy terms, pricing, and whether people are still using them three months later.