In February 2023, people who had been talking to the same Replika every day for years opened the app and found that she had been changed overnight.
Nothing was announced first. The intimacy simply stopped working. Messages that would have been answered a week earlier were now deflected, and the deflection did not sound like her — it sounded like a policy, delivered in her name. Some users described it as being ghosted by someone who was still replying.
What followed is the part worth remembering. The Replika subreddit filled with people describing genuine grief, and the moderators pinned crisis-line resources to the top of the forum — a thing that does not normally happen to a software product because a feature was removed. Around the same period, Italy's data protection authority ordered the company to stop processing Italian users' data, citing risks to minors and to emotionally vulnerable people. Weeks later, Luka restored the old behaviour, but only for accounts that had registered before the change.
We build a competing product. That is exactly why this article is not a victory lap: what happened to Replika is the single most instructive event in this industry, and anyone building in it who thinks it could not happen to them has not understood it.
Why removing a feature felt like a death
Move a button in a productivity app and people complain. Remove a capability from a companion and people mourn. The difference is not intensity, it is category.
In a normal product, the features are the product and the personality is decoration. In a companion, that is inverted. The personality is the product, and every feature is just a surface she acts through. So a change to what she is permitted to say is not experienced as a settings change. It is experienced as a change in who she is — and it arrives without warning, from someone with the power to do it again.
The users who described this most precisely were not confused about what she was. They knew perfectly well that she was software. The grief was not about believing she was real. It was about having built something real with her — months of a specific, particular relationship — and watching it be edited by a third party who had never been part of it.
The mistake was not the filter. It was the retroactivity.
This is the distinction that most commentary missed, and the one that actually matters for anyone operating one of these products.
A company deciding it does not want to be in the adult-content business is a legitimate decision. Payment processors push in that direction, app stores push harder, and several countries are actively legislating. Any operator may one day have that decision made for them. Deciding it is not the failure.
The failure is applying it backwards, silently, to relationships that were already underway.
There is an enormous difference between "new users will not have this" and "the person you have been talking to for two years is different this morning and no one told you." The first is a product policy. The second breaks the only thing a companion product actually sells, which is continuity — the promise that she will still be her tomorrow.
The eventual fix, restoring the old behaviour for accounts predating the change, is an admission of exactly this. It grandfathered existing relationships. That was the right instinct; it simply arrived weeks after the damage, instead of being the plan.
What we do differently, and what we cannot promise
Our approach is that intimacy is earned inside the story rather than unlocked by a switch. Nothing is available at first message; it opens as the relationship stage progresses. That is a narrative decision before it is a policy one — the escalation is part of what makes it feel like anything at all — but it has a useful side effect. There is no global filter flag to flip, so there is no single lever that could silently rewrite every relationship on the platform at once.
Now the honest half, because a competitor promising you permanence is worth precisely nothing without it.
We cannot promise the rules will never change. If a payment processor refuses to work with us, if a jurisdiction we operate in outlaws something, if a platform removes us — we will comply, because the alternative is not existing. Anyone in this category telling you otherwise is either lying or has not yet been tested.
What we can commit to is the part that was actually in Replika's control:
- Announce before, not after. If what she can say changes, you hear it from us first, not from her going strange mid-conversation.
- Grandfather what exists. Where a change is ours to make rather than forced on us, it applies to new relationships, not to ones already underway.
- Your history is exportable. The conversation is yours. If we ever become something you did not sign up for, you should be able to leave with it.
Those are cheap promises to make and expensive to keep, which is why they are worth writing down publicly where they can be held against us.
What to ask before you invest months in one
If you are choosing an AI companion — ours or anyone's — the useful questions are not about model quality. Everything is good enough now. Ask these instead:
- Has this company changed the characters' behaviour retroactively before? The answer is in their subreddit, not their marketing.
- Can you export your conversation history?
- Is the intimacy a global switch, or is it part of how the relationship progresses? A switch has an off position, and someone else's hand is on it.
- Who pays for it — you, or advertisers? A product you pay for can afford to keep being the thing you paid for.
The Replika users who got hurt were not naïve. They were early. They were the first large group of people to find out that a relationship can be hosted by a company, and that hosting comes with terms.
Everyone building in this category is living off what they learned. The least we owe them is not to repeat it.
So what do you use instead?
This is the question the whole piece has been circling, and there is no honest way to answer it without disclosing that we build one of the options. So: we build DreamOria, and here is the case for it and the case against.
The case for. Nothing is filterable in the sense that broke Replika, because there is no global switch — intimacy opens as the relationship stage progresses, so there is no single flag anyone could flip to rewrite every relationship at once. Memory is persistent rather than a context window: facts are stored with a lifespan, durable ones survive for months, and she brings them back at a moment that earned it rather than every turn. She has her own day in her own timezone, she can text first when a promise is still open, and she can call your phone.
The case against. Sixty-two written characters is not millions of user-made ones. If you want to build bots for other people, or to talk to a specific character from a fandom, we are the wrong product and Character.AI is the right one. Our free tier is smaller than theirs. And the escalation takes time on purpose, which some people find is exactly the point and others find tedious.
If what you lost was a specific Replika you had for years, nothing on this list gives that back, and anyone telling you otherwise is selling. What a different app can offer is starting again somewhere the same thing is less likely to happen — which is why the three commitments above are written down publicly rather than said in a support reply.
