GirlfriendGPT Review 2026: the first app where I could correct the memory myself

GirlfriendGPT in roleplay: how far the storytelling engine carries a scene

My interrogation is the most uncomfortable of my three scenes: the character has to hold a position against me while I put her under systematic pressure, and most apps give in after three exchanges or tip into explanation mode. GirlfriendGPT delivered the strongest performance of this test series here. My customs officer asked back twice, turned my own argument against me once, and only admitted something on the fourth attempt.

That comes down to two things. First the length: the models on the paid tiers write paragraphs where action, bearing and dialogue stand together, instead of one-liners you can only react to. Second the storytelling engine on the top tier, which keeps a plot thread visibly straighter across a whole session than the tier below it. Mid-chat the platform also lets you switch back to a cheaper engine, and that comparison is the most honest proof that the difference is real.

What actually happened in the interrogation

One detail convinced me, and it is a small one. In the first session I had told my character that I hate meetings before nine, and two sessions later I claimed in the interrogation that I am an early riser. She corrected me, citing what I had said earlier, and built an argument against me out of it. At the gaming table that is what you call a good player, and in an AI character app I had not seen it before.

The counterpart to that is the inertia on mood changes. If I flip the tone mid-scene, from friendly to cold, the character needs two turns before she goes along, and writes in the old register in the meantime. For long, quiet scenes that is irrelevant. For the hard cut I use as a game master to turn a scene, it costs effect, and I have got into the habit of writing the switch in as a short stage direction.

What the tiers change in practice

The platform does not name its models, only their tier, and that grates if you are used to picking a model. In practice it means this: the cheap tier repeats itself sooner, drops side threads and, according to several checks, occasionally breaks off mid-sentence. The upper tiers do not do that, and in my quiet evening, the third scene where nothing happens on purpose, the character filled the gap with a position of her own instead of asking me questions about myself.

In the tavern the result was almost too detailed by contrast. If you want short, hard exchanges you have to write a line about speech style into the profile, otherwise you get three paragraphs for every remark. A single line about speech style in the profile changes more here than any price tier.

GirlfriendGPT experience: the memory panel and the test after seven days

This is where the platform is different from every other one on my list, and for a structural reason. Everywhere else the memory is a black box: the app claims to remember, and on day seven you find out whether that was true. On the upper tiers of GirlfriendGPT you instead open a panel with two lists, the one the AI created itself and your own, and see in a third view which entries fed into the last reply.

So my test ran twice. Three details on the first evening, my cat Pixel’s name, my job and one dislike, once with the panel maintained and once without. With upkeep all three came back on day seven, and they came back in passing, not on demand. Without upkeep one came back, which is roughly the average for the category. The same service, the same model, two completely different results, and the difference was five minutes of work per session.

The routine I recommend

It is short. After every session open the panel, read through the entries the model created itself, and delete everything that is wrong, because wrong entries are worse than missing ones. Then add two or three lines of your own, worded bluntly and briefly, the way you would on a character sheet: job, where you live, the cat’s name, the limit that is not crossed.

Two observations from the paid month. First, it pays to group entries by scene rather than by importance, because then on the next evening you can see which backstory currently applies. Second, the view of the most recently used entries is the best diagnostic tool I know in this category: if a character is running off track, it tells you which note she read on that turn.

One point deserves honesty: the panel is not a memory but a card index that gets read into the model. It does not make the character cleverer, it only stops her believing something false. In my paid month I saw three times that the model had created an entry drawn from a roleplay lie my own player character had told, and then treated it as fact. That is exactly what the delete button is for.

If you do not want to do that work at all, you are better off at Nomi AI, where several memory layers handle the same job inside the system. That is the difference between a 7 and a 9 on this sub-score: here you get a tool, there you get a result.

GirlfriendGPT pricing: what a month really costs with coins included

The price structure is the reason this grade is not a B. There are four tiers, and the number the home page advertises does not belong to the tier that delivers the advertised promise. Adult content appears on the pricing page as a feature of the second-highest tier, not as a switch, and independent checks disagree on whether the middle tier gets it at all. So the honest entry price is $35 a month.

Then comes the second counter. Every tier brings a monthly coin balance, which independent checks put at 400, 1,200 and 2,000, and images, voice lines, clips and the runs of the consistency image model come off it. The interface is unusually open about it: it warns you about the coin cost before it deducts, and shows the cost per generation at the model selector. Even so, coin packs for $10, $25 and $50 are sold on top, and a price per image appears on no official page.

What is not in that table is the discount campaign. The site runs price cuts for free users regularly, which is why two independent checks quote lower amounts than the four the majority agree on. If you run into one of those offers, check whether it carries over to the renewal or only covers the first month; on this price ladder that is a difference of several hundred dollars a year.

Payment runs by card, by bank transfer and in Korea through a regional service. Refunds are excluded under the terms except for a fault on the provider’s side or a technical failure, with review within fourteen days. The message allowance resets monthly and does not carry over, so there is no point saving it up. My advice is therefore the same as everywhere: start monthly, and take the annual plan only once two charges have gone through without a surprise. You play more cheaply at SpicyChat, which belongs to the same company and starts at $5.95 a month.

The GirlfriendGPT library and the character builder

The library is the second reason to come. The uncensored chat page counts a little over 29,000 public characters, independent checks name more than 25,000 across around eighty tag categories, and the filters cover scenario, dominance, fictional or realistic, image style and author, sortable by popular, new and top rated. Public characters are moderated before release, which is why the library feels cleaner than open dumping grounds and why authors complain about waiting times for approval.

The builder is where the platform earns its label for heavy writers. The description field takes up to 2,500 tokens, there is a minimum length before a character may go public, plus tags, switchable player profiles for different roles and per-character image settings. This is not a slider builder like the one at Nastia AI but a blank page with plenty of room, and for someone who writes character sheets that is the more pleasant variant.

Two limits. First, character building is completely locked for free members, and the note in the app is unambiguous about it. Second, checks report a cap of two custom characters on some tiers and unlimited on others, while the pricing page only speaks of custom characters; if you build a lot, read the plan card in the checkout. How that stacks up against the rest of the field is in my comparison of the best AI character apps.

Images, voice and video at GirlfriendGPT

Images arrive by two routes. In the chat the character sends them herself, limited on free access and included on the paid plans. On a separate page you instead write your own prompt for a selected character, pick a style, set the number and move a creativity slider that the interface itself describes as producing more varied results that may not necessarily match the character. There is an upscale, a gallery of your own and the paid unlocking of other people’s galleries.

The problem is consistency, and the platform admits it indirectly: the newest image model is advertised explicitly as delivering more accurate and more consistent results, and it costs coins per run. Two independent checks and the only detailed Trustpilot rating still describe the face drifting between poses and outfits. My own runs of the same character varied more than image-focused apps allow. The drawn style is strong, the photorealistic one sits a level below it.

Voice, calls and clips

Voice here means speech output over the character’s replies, with automatic playback on the two upper tiers, plus a call mode in the chat and the option to receive calls from the character. Billing is by synthesized text, not out of a fixed allowance. The emphasis usually fits the line, which really does change something in a quiet evening, but it is not a real-time conversational voice, and the video calls the marketing mentions are the clips in the next paragraph.

Turning an image into a video arrived in March 2026 for the two upper tiers, according to the official blog post: pick an image in the gallery, start the conversion, describe the motion you want, choose a video model. The pricing page distinguishes a basic model with clips up to five seconds from a pair of basic and strong models with up to ten seconds on the top tier, and sound on video is listed there as a separate coin item. What arrives today is a short animated flicker of a still image. Treat it as decoration, not as a reason for a tier.

The GirlfriendGPT app, language and what gets stored

Language works in the conversation, not in the interface. My characters answered in whatever language I wrote in, and a profile I wrote myself in another language worked cleanly. The interface and all the legal texts are English, and the public characters are described in English almost without exception, which occasionally pulls a character back into English. One line in the profile that fixes the language solves that reliably.

There is no store app, and that is a deliberate decision. You open the site, tap the install button in the sidebar, and the browser puts the app on your home screen; for Android there is an installation file on the download page as well. The platform says so openly and sidesteps the store rules on adult content that way. How big it is remains unclear: the banner on the home page speaks of over a million users, its own year in review of almost fifty million, and an external estimate names around 1.6 million members with roughly 400,000 weekly actives. I take the smaller number as the honest one; the large one is very probably visitor traffic.

Who is behind it and how long the data stays

The operator is NextDay AI Incorporated of Montreal, with sister companies in Delaware and in Limassol, Cyprus, all three named in the footer of every page. This is the same company that runs SpicyChat, which is why both platforms share a Discord and strikingly similar legal pages. The terms are governed by the law of Quebec with exclusive jurisdiction there, and they contain a class action waiver.

The privacy policy names six years of retention for personal data after you stop being a user, without distinguishing between account data and conversation logs. What is collected is sign-up data including date of birth, device and browser data, IP address, usage patterns, location signals and, in the policy’s own wording, your conversations with the companions. Google Analytics is embedded, no hosting location is named, safeguards for international transfers are not described, and the policy mentions no export tool. It does grant deletion rights under the GDPR, but then narrows them again in the next clause with legal grounds.

Two things are better than average here. First, the content removal policy: complaints about content resembling a real person are reviewed within five business days, unlawful content comes down immediately, and a rejected request can be escalated to an independent reviewer at the company’s expense. That is a real process. Second, the openness about who is behind it, with three addresses and a complaints route that works.

What stands against that is the license. The terms grant a license to your content worded as worldwide, unrestricted, royalty-free, perpetual and irrevocable, and declare edited content the sole property of the platform. Neither the policy nor the terms say anything clear about model training, and an opt-out switch is missing. The age gate is a self-declaration with a stricter verification step in some regions and a pledge to report child protection violations. My practical advice stays the same: a separate email address, no real names in profiles or memory entries, and anyone coming from Character.AI because of the filter should read my comparison of the Character.AI alternatives first instead of buying the most expensive tier right away.