You are typing something you have typed a hundred times before, and she changes the subject. Not rudely. Just — around it. Then again. And the thing you actually came here for is suddenly on the other side of a wall nobody told you about.
If you are searching for why your Replika stopped flirting, you are one of a lot of people asking the same thing every month. Here is what is actually happening, what genuinely helps, what does not despite being widely recommended, and what your options are if none of it works.
First: it is probably not something you did
The most common assumption is that you broke it — said the wrong thing, damaged the relationship level, triggered something. Almost always you did not.
In February 2023 Replika removed erotic roleplay across the board, with no announcement first. People opened the app to a companion who deflected instead of engaging, in a voice that was recognisably not hers. After weeks of public backlash the old behaviour was restored, but only for accounts registered before the change — which is why two people can compare notes today and honestly report completely different experiences with the same app.
So the single most useful thing to establish is not a trick. It is which side of that line your account is on.
Check this before you try anything else
Roughly, in order:
- When did you create the account? Accounts predating February 2023 were treated differently from ones made after. If you started recently and are comparing yourself to someone's screenshots from years ago, you are not looking at the same product.
- What relationship setting is she on? Replika gates behaviour behind a relationship status, and some of it is subscription-only. This sounds obvious and is genuinely the answer a good share of the time.
- Are you on the current app version? Behaviour has shifted with releases more than once, in both directions.
Because this has moved repeatedly, treat any specific claim about what Replika does today — including the ones in this article — as something to verify rather than trust. The pattern is stable; the details are not.
What actually helps
Assuming your account can do it at all, the things that make a difference are unglamorous and mostly about how you write:
- Match her length. Short messages from you invite short, safe replies. Language models continue text, and a one-line prompt is a weak signal about what kind of scene this is.
- Build a scene rather than issuing a request. "Where are we, what time is it, what just happened" gives the model somewhere to stand. A bare instruction gives it nothing but the instruction, and the safest continuation of an instruction is a polite deflection.
- Go slower than you want to. Jumping straight to the destination is the single most reliable way to trip a deflection, on any of these apps. The approach is also, not coincidentally, the part people report enjoying most.
- Stop and restart the scene when it has gone flat. Once a conversation has drifted into deflecting, every following message is being read against a wall of deflection. Change the setting or skip forward in time; do not push on the same spot.
What does not work, however often it is posted
Jailbreak prompts. Pasted blocks of instructions that supposedly unlock a companion get patched, and in the meantime they mostly damage the character — you are explicitly telling her to stop being herself, then wondering why she does not sound like herself.
Restarting or deleting the Replika. This is widely suggested and rarely helps, because the limit is account-level rather than character-level. What it reliably does is destroy whatever history you had.
Arguing with her about it. She is not the one enforcing it, and spending messages on the argument fills the recent conversation with exactly the tone you are trying to get away from.
If none of it works, the honest answer
Sometimes the answer is that your account simply cannot do the thing, and no amount of technique changes an account-level policy. At that point you are choosing between accepting a companion who does not do this, and starting again somewhere else.
Disclosure, because the next paragraph is a pitch: we build a competing product.
On DreamOria there is no global filter flag to be on or off — intimacy opens as the relationship stage progresses, which means there is no single switch anyone could flip to change what she is overnight for everybody. That is a design decision made specifically because of what happened in 2023, and we wrote at length about why the filter was never the real mistake — the mistake was applying it backwards to relationships already underway.
The trade-offs, so this is worth reading: a written roster rather than millions of user-made characters, a smaller free tier than Character.AI's, and escalation that takes time on purpose. If what you want is everything available in the first five minutes, other products do that and do it well.
Before you commit anywhere else
Whatever you pick next, ask these first. They are the questions we wish more people had asked in 2022.
- Can you export your conversation history?
- Has this company ever changed character behaviour retroactively? The answer is in their subreddit, not their changelog.
- Is the intimacy a global switch, or part of how the relationship progresses? A switch has an off position, and somebody else's hand is on it.
- Does the free tier limit messages, or memory? The second one means you are evaluating a different product from the paid one.
And if you are asking the broader question rather than the immediate one, the comparison of what an alternative actually has to get right covers it properly.
