You have two feature tables open and they tick the same rows. Long-term memory, voice, image generation, custom personality — four ticks on the left, four on the right, and the decision has not moved. Neither app is lying. The trouble is that when you compare AI girlfriend app features from a table, you are comparing words that no two companies define the same way. A tick in the memory row can mean a list of twenty facts you edit by hand, or a system that searches two years of chat logs. Only one is worth the price you are weighing up.
Why you cannot compare AI girlfriend app features from a table
Go looking for a neutral definition of "long-term memory" and you will not find one. There is no standards body here, no agreed vocabulary, nothing obliging an app to say which mechanism sits behind the word. Almost every page defining these terms belongs to a company selling an app, and defines them in the shape its own product happens to be.
The privacy labels on a store listing are imperfect, but each store asks every developer the same questions, so the answers line up in a column. The feature table has no such discipline. It is marketing copy arranged in a grid, and the grid is what makes it feel comparable.
So read one as a list of questions, not answers. A tick tells you the app has something in that area. It does not tell you what, how much, or on which tier — and those three are the whole of the decision.
Memory is three mechanisms wearing one word
Memory decides whether a companion app is worth paying for, and it is the row where one label covers the most ground. Underneath sit at least three different machines, and they fail in completely different ways.
The context window. The model's working span — the maximum number of tokens it can use at one time. What is inside is available; what has scrolled past is gone. An app leaning on this alone feels sharp deep into a long session and blank at the start of a new one. Studies also find a model's effective context length can be smaller than the advertised one, which is this whole problem in miniature.
A written profile. A short structured set of facts — your name, your job, a few preferences — handed to the model every turn. Some apps write it as you talk, some let you edit it. Cheap, reliable and shallow: it remembers that you have a brother, not the conversation you had about him.
A searched store. The app keeps your past conversations and, when you say something, goes looking for the relevant pieces — retrieving and incorporating information from an external source rather than carrying it all in the prompt. Its failure is specific: when the search misses, the memory may as well not exist, and from the outside you cannot tell which happened.
Here the obvious conclusion breaks. Deeper is not automatically better. A store that surfaces a sad thing you mentioned in March, in the middle of an unrelated Tuesday, is worse than a window that simply forgot. And a plain profile you can correct often beats a rich one you cannot inspect. Comparing apps on memory rather than first chats goes further into what to score.
Voice is either a recording or a conversation
The voice row hides a smaller but sharper spread. It can mean the app reads its replies aloud. It can mean voice notes you play like messages. It can mean a live call where you speak, it answers, and the gap is short enough to feel like talking. It can also mean only the reverse: you dictate, it replies in text.
Four products, one word — and the distance between the first and the third is the entire reason anyone pays for voice. Latency is what you want to know and the thing no listing states. Worth checking too whether voices come from a fixed library or are generated from a clip you upload, because the second is a privacy question as much as a feature.
A tick is not an allowance
The most expensive misreading is the simplest. A tick says a feature exists. It says nothing about how much of it you get, and the metering is where the money is.
The shapes are easy to spot once you look: images sold in credits that reset monthly, so the row is ticked and the twentieth picture costs extra; voice capped by minutes or calls per day; a coin shop on top of the subscription, which makes the advertised price a floor rather than a ceiling. None of it is dishonest, and all of it is invisible in a grid of ticks.
Then the tell that costs nothing: the same word in both the free and paid column with a different adjective in front. Basic memory and advanced memory. That admits the mechanism changes between tiers — and it is exactly what a free trial cannot show you, because you will be testing the one you are not going to keep.
A ticked row tells you the app has something to say about that feature. The tier, the cap and the mechanism are three separate questions, and the table answers none of them.
Four questions that turn a tick into a fact
The honest way to compare AI girlfriend app features is to take the four rows you care about and ask each of these. Fifteen minutes, no account, no payment.
- Memory: can I see it, can I edit it, does it survive a new chat? Three yes-or-no answers identify the mechanism without the app having to name it.
- Voice: live or played back, and what is the cap? If the help pages mention minutes, credits or a daily limit, it is metered, whatever the table says.
- Images: generated for each request, or chosen from a set? And if generated, how many before it costs more.
- Which tier is this row on, and is it the same thing one tier down? The second half decides whether a trial tells you anything.
Ask them of the app's help pages, not its landing page. Support documentation has to describe limits accurately or the queue fills with people who were misled; a landing page only has to describe capability.
What this pass cannot settle
It cannot tell you quality. An app with a plain written profile, used well, feels more attentive than one with a deep store that keeps retrieving the wrong thing. Mechanism sets the ceiling; execution decides where in the room you land.
It cannot tell you taste, which is most of the decision. And it describes the app today: a tier can be re-cut, a cap lowered, a feature moved up a tier in a release you read about afterwards.
What it does is stop you paying for a row you misread, and the misreading runs the same direction nearly every time — you picture the deep version of the feature and buy the shallow one. If you would rather start from a known quantity, the app we currently recommend is a reasonable baseline, and its feature table deserves the same four questions as every other.
Frequently asked questions
What does long-term memory mean in an AI girlfriend app?
It is not a defined term. In practice it covers a model's context window, a short written profile, a searchable store of past conversations, or some combination. Ask whether you can see and edit what is stored, and whether anything survives into a brand-new chat.
Do two companion apps with the same feature list work the same way?
Usually not. There is no shared vocabulary for these features, so each company describes its own implementation, and identical rows can sit on very different systems. A feature list tells you what an app has chosen to compete on rather than what it does.
Where can you check what a feature actually does before paying?
The app's help and support pages, which have to be accurate about limits, and the in-app purchase list on its store listing, which shows what is sold separately. Between them they usually reveal the caps and tier boundaries a feature table leaves out.
Disclosure. NaughtySignal may earn a commission if you sign up through a Visit link on this page, at no cost to you. It does not change what we write. How we earn.