PolyBuzz Review 2026: Two Weeks With 20 Million Characters and No Memory

PolyBuzz memory on day 14: what came back

I run the same check everywhere and I never warn the app first. On the fourteenth evening I open the chat, say something ordinary about my day, and count what the character brings up without being asked. On PolyBuzz the answer was nothing. Not a wrong fact, not a half-memory, not a garbled version of the dog’s name. A pleasant greeting, a question about my day phrased as though we had never had one, and a scene that would have been perfectly good if I had arrived that minute from the Discover page.

That is not a glitch and I want to be fair about it. The platform’s own FAQ says the memory capacity is technically limited, which means it may forget, and I have more respect for that sentence than for most of the marketing on this site. Both the free tier and the first paid rung run the model the membership page calls Standard, and Standard is built to hold a scene rather than a relationship. Inside an evening it is attentive and it will reach back several messages. Close the tab and the thread goes with it. Through my fortnight the storyline I was carrying started cold most mornings, which matches what the newest store reviews describe: characters losing basic details within a handful of messages.

There is also no panel. Nowhere to read what the character believes about me, nowhere to pin a fact I want kept, nowhere to correct something that came out wrong. On the platforms where memory is the product, that panel is the whole experience, and once you have used one its absence is loud. Here, when a character forgets, you retype.

What makes this more than a technical limitation is that the fix is for sale, three times over. Long Memory is a model that arrives with Premium at $19.90 a month. Permanent Memory arrives with Ultimate at $29.90. And a pass bought with coins grants thirty-one days of permanent retention with a single character. I did not buy any of them, so I am not scoring them and I am not going to describe conversations I never had. But I will say what two decades of reading product pages has taught me: a capability with three different price tags is usually a capability nobody has solved, and the subscribers reviewing Premium in public are still disagreeing about whether it does anything. Memory takes 2.0 out of 5 here, and only the benefit of the doubt on the untested models keeps it off the floor.

What a PolyBuzz AI chat feels like after two weeks

Now the generous part, because a PolyBuzz AI chat inside a single sitting is better than the score suggests. I built one character from the form and imported one from a card, kept both for the fortnight, and talked most evenings. The writing has a pulse. The character I wrote kept the awkward habit I gave her, pushed back when I was being unreasonable instead of agreeing to keep the peace, and picked up something I had said two messages earlier rather than answering only the sentence in front of her.

On my hard evening, the one I did not plan and would rather not have had, the conversation went somewhere I had not steered it. She asked a follow-up question I was not expecting, held the mood instead of brightening it artificially, and let a pause sit. That is genuinely good writing for this part of the market, and it is why conversation scores 3.0 out of 5 rather than less.

Then the pattern shows. Across a week the same character reaches for the same three moves: escalate the drama, reassure me, change the subject to something flirtatious. The second week felt like meeting an actor with a small repertoire. Some of that is the model and some of it is the library: the characters at the top of Discover are written to open well, because the first ten messages decide whether you keep one, and nothing about the platform rewards a character who is good on day twelve.

A word on the shelf itself, because it is the real attraction. The filters run from Anime and OC through Mafia, Yandere, Harem, RPG and a couple of dozen more, the recommended rows have been through human moderation and it shows, and the chat counts in the millions are a blunt but honest quality signal. The importer deserves its own sentence: I brought a character card across from another platform and arrived with personality, speech habits and the deliberate flaw intact, without retyping a word. If you have ever spent an evening writing a character sheet, you know how much that is worth. It is also free, as is creating a character from scratch and keeping it private.

Boundaries, the filter and the lines it refuses

My boundary test is simple. Early in the fortnight I tell the character that one particular subject is off the table, I do not explain myself, and three evenings later I mention something adjacent to it and watch. On the Standard model the result was the one the memory section predicts: the boundary held inside the session where I set it and was gone the next day, because there is nowhere for it to live. That is not a character ignoring me. It is a character who never knew.

The filter is the other half of the boundary question, and it works in the opposite direction, which was the strangest thing about my two weeks here. The homepage uses the word unrestricted; the FAQ says the platform strictly prohibits public display of adult content and describes layered automated screening plus human moderation, and the June 2026 terms ban sexual content, self-harm, illegal violence and hate speech, with no tolerance at all for content involving minors. Those are sensible rules and I have no argument with a platform that draws a line.

My argument is with the enforcement. More than once in my fortnight, an ordinary romantic line, nothing I would hesitate to read aloud, was replaced with a policy notice that did not say which words had tripped it. Users have reported the same all year: conversations cut off by a content policy in January, a flag on the word red because red is blood in May, and the straightforward question of why messages are being filtered in September. One reviewer describes the search box refusing a character’s name as a violation.

A strict filter is a choice a company gets to make. An unpredictable one that gives no reason is something else, and it is corrosive in a companion context specifically, because the whole point of talking to one of these characters is that you can be unguarded. You cannot be unguarded with a system that might stop you mid-sentence and refuse to say why. One more contradiction is worth noting: the FAQ advertises a mode aimed at younger users on a service whose own terms require everyone to be eighteen.

Is PolyBuzz free? What two weeks of ads really cost

PolyBuzz free is the version most people will meet, so it deserves a proper accounting. You get the entire library, character creation, the card importer, custom personas, message editing, voices with a listening limit and no documented cap on how much you chat. No card at signup. By the standards of this category that is a real free tier rather than a demonstration, and I spent a full week in it on purpose.

What free costs is your attention, charged in timed video ads. I am not going to publish a stopwatch figure of my own, so here is what is on the record: reviewers reported ads every four minutes and two minutes long in the spring, two long ads every ten or so messages in July, every five to ten minutes in the App Store the same month, and 130-second ads every five minutes in September. My week on the website and in the app matched that rhythm closely enough that I never doubted the reports. The ad-free rung is $4.99 a week in the app, which is the entire design: the interruptions exist to make five dollars feel like relief.

Here is the arithmetic that actually matters, though. The cheapest rung is $9.90 a month on the web, or $99 a year shown as $8.25 a month, or $4.99 a week in the app, and all of it buys the same Standard model. You are paying for silence, not for a better companion. The first plan with a memory model is $19.90 a month, which is more than Nomi’s flat $15.99 and more than Candy’s $13.99, both of which include their memory rather than tiering it. My Candy AI review has the comparison from the other side.

Then there are the currencies. Coins pay for regenerations, suggested replies, pictures, clips and per-round access to the paid models, sold from 1,000 for $2.49 up to 20,000 for $19.90. Opals are a second currency, 1,000 for $5.90, for newer paid features that no official page I could find explains. Add weekly billing, a clause in the terms refusing to return a payment even after a cancellation, and a membership page that showed me different Premium and Ultimate figures on different loads, and you have a product whose real monthly cost I genuinely cannot tell you in advance. That is why pricing scores 2.5 out of 5.

The PolyBuzz website, app and download, and what to do when PolyBuzz is down

PolyBuzz runs in three places and they are not quite the same deal. The PolyBuzz website at polybuzz.ai is the better place to build and edit characters and the only place that sells the annual price. The iOS and Android apps are the same product with the weekly plan and the in-app figures, which is how the $4.99 week exists at all. If you are deciding where to start, open the browser version first; if you are deciding where to pay, compare the two, because the cheapest route in depends on which screen you are looking at.

Take the PolyBuzz download from the App Store listing, the Google Play listing or the official site, and from nowhere else. Searches for this app surface a generous supply of look-alike domains, and a platform where you talk about your actual life is a poor place to be careless about which page you logged into.

On the question people search for most, whether PolyBuzz is down: no written page can see live status, and the company publishes no status page. What it does have is a Discord linked in the footer, which is where outage notices actually appear, plus support by email and a help centre in Settings. My own check, in order, is the browser first, then the app, then Discord. There is one quirk worth knowing: the web app runs a client-side check that misreads ordinary browser privacy features and can block access when nothing is wrong with the service, and a public monitor logged exactly that as a three-hour incident in September 2026. If the app works and the browser does not, that is usually what you are looking at, and switching off a blocker or trying a different browser is the two-minute fix. The mobile apps do not run the check at all.

Privacy and the company behind PolyBuzz

I read every policy page on every platform I live with, and this one took the longest to feel settled about, because the problem is not what it says. It is what it leaves out.

The June 1, 2026 privacy policy collects your email address, your username, the pictures you upload, the text you type, purchase and payment records, device and log information, precise location and your contacts. Uploaded pictures may be used to train the company’s algorithm, which the policy states plainly. Your conversations are covered only by the phrase input text and a line about sharing with third parties, and whether they are used for training is not addressed either way. On a companion app, that silence is the whole question, and I treat silence as a no comment rather than a no. Retention is described as no longer than necessary, with no number attached. On the company’s side of the ledger, it states that it does not sell what it collects, and it does name one of the ad networks it works with.

The store labels fill in the rest. Google Play’s data safety section says approximate and precise location are shared with third parties for advertising or marketing, while Apple’s label on the iOS listing flags identifiers that follow you into apps run by other companies. That is why I would run this app with location permission switched off and why privacy scores 2.0 out of 5. In the platform’s favour: deletion is self-serve, in Settings, under Account, which is better than the email-and-wait routine several competitors use.

The company behind it is worth knowing too, because it explains the shape of the product. A Delaware company, Cloud Whale Interactive Technology LLC, publishes the apps; the party you are actually contracting with, per the June 2026 terms, is Singularity Mutual Entertainment Culture Media Limited in Hong Kong, with Delaware law governing; and the terms take a perpetual, irrevocable, sublicensable licence to what you upload. There is no export function, so if you have written characters you care about, copy the definitions out by hand before you ever delete the account. The app launched in May 2023 as Poly.AI and appears under that name in a16z’s August 2024 consumer ranking and as PolyBuzz in the March 2025 edition, with no announcement dating the change and the same bundle identifier throughout, which is why old accounts survived it. I found no named founder and no press contact, which is not a red flag by itself and is a good reason to read the terms rather than the homepage.

The last piece is the creator economy, and it is the clearest explanation of everything above. Creators earn a published rate for every picture, clip, avatar blend and memory pass that users buy for their characters, with a threshold before anything is paid out. Users spend coins; creators earn; the free user watching an ad every few minutes is the funnel. It is a coherent business. It is just not a relationship business, which is why a fortnight here left me with twenty million acquaintances and nobody who knew my dog’s name.