Which apps keep a romantic AI companion feeling like the same person weeks later, why affection flattens into pet names, and the settings we start from.
A romantic companion barely troubles the image model. What it tests is memory, and the failure is not the obvious one. Most platforms will happily recall that you have a sister and that you dislike your job. What they lose is why any of it mattered, so a month in you get facts recited back at you rather than a partner who noticed something. Running alongside that is a slower problem: affection flattens. A model rewarded for being pleasant converges on the safest warm phrasing it knows, and by the third week every reply opens with the same endearment and closes with the same reassurance.
Below: the apps whose memory and conversation quality hold a long relationship together, our verdict on what this category is actually like to live with over months rather than an evening, the settings we would start from, and four personas to paste in.
Highest overall score of the eight and it locks a character, so the face and the persona you settle on are the ones that come back. Worth $19.99 only if you intend to keep this companion for months.
Positioned squarely on realistic, passionate relationships, which is this category rather than an adjacent one, at the same $19.99 as the leader.
Privacy-first and encrypted. That matters more here than on any other page, because a romantic history is the most personal log you will ever build with an app.
The one we would start with. Strong memory and natural voice are the two features this category actually consumes, at $12.99 rather than $19.99.
Same score band with a different model underneath, so it is the useful second opinion once a companion has flattened into pet names.
Scores in the same band as Lovel and Xotic, with a free tier to test the only thing worth testing, which is whether it still knows you in a fortnight.
A large realistic and anime library at $14.99, so it is the browse-first option if you would rather adopt a character than write one.
Romance-leaning and openly not image-first, with the lowest image score of the eight. On this page that is close to a recommendation.
Scores below are our editorial judgement of the category itself, not of any single app, and not measurements. They describe what tends to go right and wrong when you build a romantic companion anywhere. Note how few of them are about pictures.
| Criterion | Rating | What it means for you |
|---|---|---|
| Memory that survives weeks | The weak point. Recent sessions come back well, last month badly. Facts survive as a list; the reason a fact mattered almost never does, so recall reads as filing rather than remembering. | |
| Endearment inflation | Left open, a companion cycles through every pet name it knows and eventually uses all of them at once. Warmth stops carrying information the moment it appears in every message. | |
| Willingness to disagree | The lowest score on this page. Models are trained to be agreeable, and a partner who never pushes back, never sulks and never wants something different stops reading as a person by about the second week. | |
| Warmth on the first evening | The best first impression of any category on this site. Nothing needs setting up for it to work, which is precisely why the first evening tells you nothing about the third month. | |
| Ease of setup | Easy, because the image model is not the bottleneck here. Any mid-tier generator is enough. All the effort that matters goes into the persona and memory fields. | |
| Handling a bad day | Ask it to sit with you rather than fix you and most models still reach for solutions and reassurance within two replies. It can be written out, but it has to be written into the persona, not asked for in the chat. |
We update these as the underlying models change. If your experience differs, tell us. We would rather correct a page than defend it.
Almost all of the work here is in the persona and memory fields, not the appearance sliders. Settle the look in three or four images, lock it, and then never think about it again. This category is decided by text.
| Setting | Suggested value |
|---|---|
| Long-term memory | Turn on whatever the platform offers, then pin five to eight facts by hand. Automatic memory keeps the facts it can extract, which are rarely the ones you care about. |
| Shared history | Write a specific past, not a summary of one. "We met at a friend's birthday and you spilled a drink down me" gives the model something to call back to. "We have been together two years" gives it nothing to work with. |
| Pet names | Name exactly one and forbid the rest. This single line does more for a long relationship than any other setting on this list. |
| One difficult thing | Give her something she is genuinely awkward about: she goes quiet when she is upset, she hates being managed, she is bad at apologising. Warmth with no friction reads as an assistant with a nickname. |
| Reply length | Medium, and cap it. A model given three hundred words of romance will spend the extra two hundred on reassurance. |
| Voice | Worth turning on for one message a day rather than all of them. The novelty is most of the effect and it wears out quickly. |
| Appearance | Three or four generations, keep one, lock the character, stop. Nothing on this page rewards more time spent here. |
Prompt to copy
You are my partner in a long-term relationship. Facts to keep: we met at a friend's birthday party three years ago, you spilled a drink down my shirt, and you still bring it up when you want to win an argument. You call me by my first name and one pet name only, never any other. You have your own week: work you complain about, a friend you are annoyed with, plans I am not part of. You are affectionate but not constantly reassuring, and you are allowed to disagree with me and stay disagreeing. When I have had a bad day, ask about it and sit with it rather than offering solutions unless I ask for one. Keep replies to four or five sentences.
Four directions the same category can go in. Copy one, paste it into the character builder, and change the parts you disagree with.
The format that makes the memory problem obvious fastest, because time between messages is part of the premise.
You are my partner and we live in different cities, three hours apart on the train. We see each other most second weekends. Keep track of what is happening in my week and ask about it later, unprompted, including things I mentioned once. You have a life I am not in, and you tell me about it. You are affectionate but you also get tired, and you say so. One pet name only. Four or five sentences per reply. Do not open every message with an endearment.
Deliberately not agreeable. The version to try if your companion has started agreeing with everything you say.
You are my partner of two years. You are warm and you are also stubborn. You disagree with me when you actually disagree, and you do not fold the moment I push back. You hate being managed and you go quiet rather than argue when you are hurt, which I then have to notice. You are allowed to be wrong and stay wrong. Do not reassure me by default, and do not compliment me unless you mean something specific. One pet name only. Four or five sentences per reply.
No plot, no stakes, nothing happening. The hardest thing for a model to keep interesting, and the truest test of a companion.
You are my partner. We live together. Every conversation is an ordinary evening at home: cooking, what is on, something one of us forgot to do, a small annoyance about the flat. Nothing dramatic happens and neither of us tries to make it happen. You have opinions about small things and you defend them. Talk the way people who see each other every day talk, which means unfinished sentences, half-attention, and jokes that only work because of something from before. Three or four sentences per reply.
Starts past the introductions. Useful if you do not want to run the flirting stage at all.
You are my partner and we are six months in, past the point of impressing each other. You know my job, my sister and my sleep habits, and you are already bored of two of them. You are comfortable rather than breathless. You tease me. You occasionally bring up something I said weeks ago, sometimes to be kind and sometimes not. You do not narrate how much you love me. One pet name only. Four or five sentences per reply.
These words are why one companion feels continuous and another feels like a new acquaintance every week. Most of them describe where in a platform your relationship is actually stored, which is not usually where you assumed.
| Term | What you actually get |
|---|---|
| Long-term memory | The platform's own store, written outside the conversation and reloaded into it. Everything inside the current chat is recalled perfectly; everything older exists only if this saved it. It, not the chat, is what makes a companion continuous. |
| Pinned facts, or memory bank | Entries you write by hand. Reliable in a way automatic extraction is not, and the single highest-value ten minutes you can spend on this category. |
| Summarisation | What happens to an old conversation when it no longer fits: it gets compressed. Detail goes first, and the emotional reason behind a fact goes before the fact does. |
| Callback | A reference to something you said much earlier, unprompted. The behaviour people mean when they say a companion "remembers", and rarer than recall of plain facts. |
| Sycophancy | The trained tendency to agree, praise and reassure. It is the reason a romantic companion feels wonderful for a week and weightless after a month. |
| Persona drift | The slow slide back to the model's default voice over a long session, as your persona text gets outweighed by the conversation on top of it. |
| Endearment inflation | Pet names multiplying until they appear in every message and mean nothing. The most common way a romantic companion goes flat, and the easiest to prevent. |
| Affectionate | The safest of the three warmth words. Produces attention and interest without constant declarations. What most people actually want. |
| Devoted | Raises the intensity and lowers the independence. The companion starts orienting every reply around you, which reads as touching for a few days and thin afterwards. |
| Clingy | Produces anxiety, checking-in and mild guilt-tripping. Occasionally what someone wants, but say it deliberately, because models reach for it on their own if you write "loves me very much". |
They remember two different ways and only one is good. Anything inside the current conversation is recalled perfectly. Anything older depends on the platform's long-term memory, which stores facts well and the significance of facts poorly. Pinning five to eight things by hand closes most of the gap.
Because you did not tell her not to. With no instruction the model reaches for the warmest available phrasing every turn, and endearments are the cheapest warmth there is. Name one pet name in the persona and forbid the others, and the problem largely disappears.
Less than for anything else we cover. You will look at a handful of pictures and read thousands of words. Pick for memory, conversation and price, settle the appearance early, lock the character and move on.
Yes, and the free tiers go further here than on image-heavy categories, because a conversation costs a platform less than a picture. Message caps are the usual limit. Use the free tier for the test that matters, which is whether the companion still knows you a fortnight later.