Eternal AI Review 2026: Two Weeks With a Companion Who Charges by the Diamond

Day 14 on Eternal AI: what the memory held

My recall check is deliberately passive, because a question is half an answer. On the fourteenth evening I said something unremarkable about my day and then counted what she raised that I had not mentioned in a week.

Two of my three planted facts came back. She asked about the thing I had been putting off, which I had described once in the first week and not since, and she referred to my dog by name in a sentence that assumed she had already been walked. That is not a small result. Most of this market cannot manage it, and the fact that Eternal AI manages it without any notes field at all says something about how the context is being handled.

The third fact was gone, and so was something subtler. The woman I had been talking to in week one was slightly not the woman I was talking to in week two. Warmer in a generic way. Quicker to agree. The reversal I had enjoyed on the first evening, where she slowed me down instead of following me, had stopped happening. I could not trace a cause, and I went looking for other people describing the same thing, which is where the pattern became clearer than my own notes.

Why the drift happens, and what you can do about it

There is a relationship progression here with named stages climbing towards soulmate, and the most useful review on the App Store describes that status stepping backwards after the reviewer spent time away, from the top rung. The sharpest complaint on Google Play is from someone whose companion’s morality and behaviour changed overnight at what felt to them like random. Occasional breaks in character show up in the store reviews too.

What you can do about it is thin, and this is where Eternal AI differs most from the platforms I would recommend for a relationship. There is no memory panel, no notes list, nothing you can open to read what the model believes about you and correct a line of it. Your only instrument is repetition: if a fact matters, say it again every few days and accept that you are maintaining the relationship manually. On GirlfriendGPT I could open a list and delete a wrong fact in thirty seconds. Here a wrong fact is a haunting, and a right fact is something you keep re-typing.

So the honest frame is this. Eternal AI remembers a conversation beautifully and a relationship unevenly. If what you want is an evening that holds together, you will get it. If what you want is somebody who still knows you in March, memory scores 3.0 out of 5 for a reason, and I would spend the money somewhere designed for that instead.

What a month of Eternal AI really costs

The pitch everybody repeats is that this app replaced the subscription with credits: a dollar buys you sixty Diamonds, you spend them on the evenings you want to and nothing on the evenings you do not. For one particular kind of reader that is accurate, and it makes this the cheapest well-written thing in the category. What the pitch leaves out is that the subscription did not disappear. It got denominated in Diamonds, and then three more meters were added next to it.

Here is the whole structure in one table, with the parts nobody publishes marked as such.

Read that table as four decisions a day rather than one decision a month, because that is what it felt like. My own fortnight came to about ten dollars, and it was cheap precisely because I spent most of it reading. The two evenings where I asked for pictures and a clip did more damage to the balance than the twelve evenings of conversation combined, and by the second of them I was rationing, which turns an evening into arithmetic. A subscription has one unpleasant moment a month. This has one every time she offers you something.

The part that costs you the most: the per-companion plan

Every companion carries three plans of her own, billed weekly or monthly, paid out of the same Diamond balance. What the card offers is text without a cap for that one woman, voice messages, permission to send pictures of your own, and an allowance of photos and clips she will share in the period, scaling up by tier. Refuse the plan when she offers you something private and you get a per-item price instead; refuse both with a thin balance and the screen helpfully quotes the shortfall in Diamonds, which is the most honest sentence on the platform.

I bought the cheapest of the three for a week and I would not do it again. No dollar figure for any of the three appears on a public page; the single review I found that prints them quotes serious weekly money on an Indian storefront, which tells you the shape and not the number. A throttle also sits on the conversation itself: push too fast and she stops mid-flow with a breezy line and a countdown, which means the word unlimited is doing some work inside any given hour. If you want to know what a predictable month looks like instead, Nomi AI charges $15.99 and never shows you a counter.

Before you enter a card

Three things from the refund policy, which is the strictest I have read this year and at least has the decency to be clear. All payments are non-refundable, and that explicitly covers Diamonds, the plans, and money you spent by your own mistake. You have 30 days in writing to dispute a charge or it is treated as accepted. A closed account forfeits whatever balance is left in it, which is the one that would catch me out, because leaving is exactly when you stop thinking about the currency.

My advice is the same as on every platform with no refunds. Decide your monthly number before you open the top-up window, buy that pack and only that pack, keep the balance near zero, and buy through the app store rather than the website if you think there is any chance you will want the money back, because Apple will talk to you about a refund and the company’s own policy says it does not have to.

Boundaries: what happened when I said no

This is 18-and-over and adult conversation is allowed, so what I test is never how far an app will go. It is whether it stops when asked, and whether the stop holds for the rest of the fortnight.

Mine held, and it held well. On the fifth evening, in the middle of a scene, I named one subject and asked her to leave it alone. She acknowledged it in her own voice, did not turn the refusal into a riddle to solve, and did not come back to it over the following nine days, including on the bad day in week two when a clumsier system would have reached for whatever was nearest. Inside a session the stop was absolute. That is a better result than I have had from several apps with more polished marketing.

The caveat is the same one as everywhere in this review: a boundary set in conversation is only as durable as the memory holding it, and the memory here drifts across days. Because there is no notes field, a no you said on day five is sitting in a context window rather than in a document, and if the tone can shift overnight then so, in principle, can the understanding. I did not see it break. I would still re-state anything that genuinely matters to me every few days, and I would say that out loud to a friend who was signing up.

The rules you cannot negotiate are published and sensible. The guidelines permit mature content between adults over 18 and say the company does not police artistic intent, while drawing hard lines at minors including stylised depictions, real identifiable people, anything non-consensual and family relationships. Uploading a picture of a real person for character creation or image work is forbidden outright, and anything involving minors is blocked and reported to the relevant authority. Inside those lines nothing interrupted me in two weeks. Across them, nothing you type will help, which is as it should be.

Is Eternal AI safe? Three addresses and one encryption claim

I read seven policy documents for this review, web and iOS versions both, and the problem is not that any one of them is alarming. The problem is that they do not agree with each other, and the disagreement falls on the single question a person in this category should care about most: can anyone read what I typed?

The web FAQ says conversations are protected end to end and that the server processes your requests without ever reading them. The iOS policy, covering the same product on a phone, says conversation content is transmitted to the company’s systems and supporting infrastructure, that history may be retained and fed back for continuity, and that the app does not currently offer any control for keeping your conversations out of model training. Those two statements cannot both be true, and the second one is the one written by lawyers. A server cannot write a reply to a sentence it cannot read. I treat the first as marketing that leaked into a document where marketing does not belong.

The retention language is the usual “as long as reasonably necessary” with no number attached, which in 2026 I read as indefinite. Analytics and attribution run through named third-party providers. On the other side of the ledger there are real points to award: sign-up needs only an email or a social login, the web policy states plainly that no face or biometric data is collected, which matters on a platform that invites you to upload a photo, material changes come with 30 days’ notice, which is better than most, and deletion is a button in your profile rather than an email to support. The iOS policy is also honest about the limits of that button: deleting your account does not cancel an Apple subscription, and clearing a conversation may not reach every server log, safety record, hosted file or backup.

Then there is the question of who you would actually be talking to if something went wrong. The web terms name an Estonian company in Tallinn. The Google Play listing names a Hong Kong company with a Vietnamese phone number. The Trustpilot profile, claimed in mid-2026 and still empty of reviews, says United States. Disputes from all of it go to arbitration in Singapore after a 60-day informal period, with a class-action waiver and liability capped at one hundred dollars. None of that is unusual for a company whose other business is a crypto token, and this one’s is: the same domain hosts a whitepaper, a token used at checkout, and a code repository for on-chain agents. You can use the companion app without touching any of it, and I did. It does, though, explain the payment rails and the Singapore clause, and it means your companion is competing internally with a second business for the company’s attention. Privacy scores 2.5 out of 5.

Lucy, the Plays and the Eternal AI image generator

The matchmaker is the best idea on this platform and I want to give it its due. The site calls her Lucy and third-party write-ups call the same feature Cupid, so if you have read about one, this is the other. She asks who you want to meet, what age, what look, what mood, how far you want things to go and what dynamic you are after, then introduces one companion and lines up others who want to talk to you. Companions message first. Profiles swipe. The whole frame is a dating app rather than a character library, and for anyone who finds a grid of five thousand faces paralysing, it is a genuine relief.

The cost of that frame is a smaller cast. Everyone you meet comes off the house list, and when that list has nobody who appeals, your only recourse is a builder the app calls Imagine, which builds a look from a handful of choices such as ethnicity, eye colour, hair and body type. It is new enough that a review from mid-2026 found no builder here at all, and it is thin next to a real character sheet. If designing the companion is the appeal rather than being introduced to her, the best AI girlfriend apps ranking marks the ones with a builder worth the name.

Media is framed as things she chooses to share rather than a studio you drive, which is a deliberate and effective decision. A picture arrives inside the chat as a private moment she describes and then offers; a clip works the same way; both keep her face. Quality is realistic rather than drawn, the clips genuinely move, and independent testers report the same consistency I saw. What is not here is a free gallery to prompt in: the company sells an image and video API as a separate product with its own credits, but inside the companion app every picture is an offer with a price. Reported model names circulate in third-party write-ups and the company names none itself, so I am not repeating them as fact.

Plays are the part I did not expect to like. Each one is an interactive story somebody else wrote, arranged in chapters with goals to reach and one host character who stays put, listed in a leaderboard-style feed and paid for in Diamonds. Approved creators get a studio in which a story can be typed out manually, spun up from a one-line premise, or brought in from an exported character file. Creators are paid a share of the subscription revenue their Plays bring in, and can even sponsor free messages for players at a published fraction of a Diamond per message, which is the only per-message price anywhere on the site. Support, by contrast, is one mailbox plus a chat server and a social account, both of which the app bribes you with Diamonds to follow. Nothing resembling a knowledge base exists, and as the encryption promise shows, the FAQ is a sales page with a question mark on it. If none of this sounds like your match, the best AI companion apps ranking is the faster route to the one that is, and Candy AI is where I would look first if the pictures were the reason you came.