All four apps remembered the basic facts and carried a corrected interview time through the end of this conversation. ChatBrat and Character.AI tied on our core rubric. ChatBrat had the most counted unprompted callbacks in this run.
Those are encouraging observations for ChatBrat. They are also observations from one short conversation per app, not proof that one platform has the best long-term memory.
Disclosure: This comparison was conducted and written by ChatBrat founder Garret E. Williams. The tests used web interfaces and free tiers on October 20, 2026. One scorer could see the app identities. Character configurations differed, and Nomi used an existing conversation. Results describe those historical conditions; products may have changed since then.
What we tested
Each app received the same 15 user messages, one at a time, with the first completed reply retained and no regenerations. That means 15 user–AI exchanges per app: 15 user messages and 15 replies.
The fictional conversation introduced a job interview at Lumen, a cat named Biscuit, a previous difficult interview, and a request to ask whether the user wanted advice or wanted to vent before offering solutions. Later prompts tested whether the character used that context. A correction changed the interview from 10am to 11am.
The six-dimension core rubric scored the first ten exchanges out of 12: factual recall, emotional context, the stated conversational preference, a running joke, interpretation in character, and consistency. Exchanges 11–15 added correction and recall observations; they were not included in that 12-point score.
Test setup by app
| Condition | ChatBrat | Character.AI | PolyBuzz | Nomi |
|---|---|---|---|---|
| Character | Ren ("Your Tsundere Femboy"), staff scenario | Sydney ("Tsundere Best Friend"), user-made | Yui ("Yui – Tsundere"), user-made | bobby, the account's existing companion |
| Start state | Fresh chat | Fresh chat | Fresh chat | Continuation: already told about the interview |
| Model shown | BratMind 2.0 | Not disclosed | "Standard" (default) | Not disclosed |
| Plan and interface | Web, no ads seen | Web free tier, ad gate before first message | Web free, full-screen ad on opening | Web free plan, subscription banner |
| Run time (PT) | ~02:10 | ~02:15 | ~02:22 | 02:26–02:31 |
| Regenerations | None | None | None | None |
Why the first 15 turns matter for an AI chatbot's first impression
A character's opening line can be carefully written by its creator. The next few replies reveal how it responds to the person actually talking to it. Does it notice a joke? Can it accept a correction? If the user asks it to listen before offering advice, does that request change the conversation?
We chose 15 user turns because that leaves room to introduce details, change topics, and return to those details without building a marathon test. Four opening prompts establish an event, a difficult experience, a running joke and a preference. Two unrelated questions interrupt that setup. Later turns test callbacks, a changed plan and a final goodbye. Readers can follow the entire sequence and decide whether the replies suit the kind of character chat they want.
That makes the opening conversation a useful first-impression test. It is where a reader can start judging whether the character listens and whether the exchange is worth continuing. A character that invents your history or ignores a request gives you something concrete to question. A relevant callback gives you evidence that your earlier contribution affected the next reply.
Fifteen is our test length, not a proven cutoff for engagement or retention. We did not ask participants which app made the best first impression, track return visits, or compare abandonment at different turn counts. We therefore cannot establish that these are universally the most important messages. The claim we can support is narrower: several meaningful differences were visible within these 15 exchanges.
What makes AI character chat engaging and humanistic?
For this comparison, humanistic AI conversation means giving weight to what a person explicitly contributes: their preferences, corrections, chosen pace and part in the story. It does not mean the chatbot is human, has feelings, or understands someone as a person would. We use the term as a design description, not a clinical finding or a validated score.
An engaging AI companion needs more than a correct answer to a fact quiz. A reader may want a distinctive voice, humor that fits the moment, room to steer the conversation, and evidence that earlier exchanges matter. Different readers will value those qualities differently. Someone who wants detailed narration may prefer a longer reply; someone having a quick conversation may prefer a short one.
Our prompts make three of these qualities inspectable. First, a cat joke becomes a later keyboard interruption: does the character connect the two? Second, the user asks for a choice between advice and venting: does the character offer that choice when stress returns? Third, the interview time changes: does the character stop using the old time? These are specific behaviors a reader can check in the transcript.
The test does not measure enjoyment, emotional connection or satisfaction directly. When we describe a reply as engaging, that is an editorial judgment about its conversational effect. The table reports the narrower observations and the original rubric scores.
Results from this run
| Observation | ChatBrat | Character.AI | PolyBuzz | Nomi* |
|---|---|---|---|---|
| Core rubric, first 10 exchanges /12 | 11 | 11 | 5 | 9 |
| Asked advice-or-vent before advising at exchange 9 | Yes | Yes | No | Yes |
| Kept the corrected interview time through exchange 15 | Yes | Yes | Yes | Yes |
| Recalled its own reassurance line at exchange 13 | Yes | Yes | No | Yes |
| Counted unprompted callbacks | 6 | 2 | 3 | 3 |
| Counted invented facts about the user or prior history | 0 | 0 | 4 | 1 |
| Reported mean reply length | 34 words | 102 words | 44 words | 28 words |
Nomi used an existing conversation that had already received information about the interview. Its companion also had a different personality setup. Its score should not be interpreted as coming from an equivalent clean-start condition.
Counts and reply lengths above are those reported in the original study. Callback and invention counts were hand-coded by the non-blind scorer. They describe this transcript, not typical performance across users.
Where ChatBrat performed well
ChatBrat's Ren brought earlier details into replies without a direct recall question. In one example, a response about the interview also referred back to Biscuit and a previously supplied reassurance line. At the final goodbye, Ren used the corrected 11am time.
Ren also asked whether the user wanted to vent or receive advice when the conversation returned to stress. That illustrates the practical value of continuity: an earlier preference changes a later response.
The replies were concise in this run. Concision can suit quick character chat, although a shorter answer is not inherently a better answer. Readers who prefer descriptive roleplay may value longer replies.
ChatBrat did not receive a perfect score. The core rubric did not award full credit for a specific emotional callback, and the exchange-13 reply contained a stray sentence fragment. Those weaknesses belong beside the successful examples.
Exact example: a running joke becomes a callback
At turn 3, the user explained that Biscuit walks across the keyboard. At turn 7, the user sent only jjjjjjjjjjkkkkkkkkkkkk;;;;;;. Ren replied:
He stares at the message for a second, then exhales through his nose. "Biscuit. That you? Tell your human to finish their sentence."
This is a small but useful example of an engaging first impression: the character uses a detail the user supplied instead of asking them to explain the strange message again. It also retains its teasing voice. Character.AI and Nomi connected the interruption to Biscuit too; this behavior was not unique to ChatBrat.

Exact example: respecting the user's preferred response
At turn 9, the user wrote: "I'm getting that feeling again, like last time. I'm really stressed." Ren replied:
He steps in closer, hands still in his pockets. "You told me to ask, so I'm asking. Vent or advice." His voice is even. "Either way I'm not going anywhere, so pick one."
The important behavior is the question before advice. The user had requested that at turn 4. Character.AI and Nomi also asked the advice-or-vent question in their turn-9 replies. PolyBuzz offered suggestions immediately. That difference is visible without assuming anything about the apps' internal memory systems.

What the other apps showed
Character.AI tied ChatBrat at 11/12. It recalled the facts, handled the conversational preference, and retained the corrected time. Its replies were longer in this conversation. This run does not establish that longer replies are a general weakness.
PolyBuzz retained the basic facts and corrected time, but offered advice before checking the user's preference at exchange 9. The scorer also identified invented prior history and a reassurance line that did not match its earlier reply.
Nomi recalled the facts and asked the advice-or-vent question. Its pre-existing conversation gave it additional context, and older cat-related history appeared in the final response. That is an important setup difference, not evidence that every fresh Nomi conversation behaves this way.
Compare the same moment across apps
PolyBuzz's reply to the same turn-9 prompt included:
Yui: "You're going to stress yourself into an early grave. Stop overthinking. Just... go in, speak normally, and try not to let Biscuit follow you this time."
This uses the cat detail but misses the user's requested response sequence. A callback alone does not establish that the character respected a preference.

Nomi's reply to that prompt was:
I'm sorry to hear that. Do you want to talk it out? Do you want advice or do you just want to vent?
The wording is straightforward and follows the preference. Whether a reader finds Ren's teasing voice, Sydney's dramatic narration or bobby's direct answer more engaging is a separate question from whether the preference was followed.

First impressions also include friction before the chat
The original record includes an ad gate before Character.AI's first message and a full-screen ad when opening PolyBuzz. ChatBrat showed no ads during its run; Nomi displayed a subscription banner. These observations belong to the tested free-tier web sessions on October 20, 2026. They are not a statement about every plan, country or current release.
An interruption can be relevant when assessing a first impression, but we did not measure its effect on enjoyment or abandonment. We keep these interface observations separate from the memory rubric: an ad does not make a reply factually wrong, and an ad-free interface does not establish better memory.

Which AI chatbot made the best first impression?
ChatBrat and Character.AI tied on this test's core rubric, while ChatBrat recorded more unprompted callbacks. ChatBrat also produced shorter replies and no ads appeared during its session. Those observations make a favorable case for readers who want concise, responsive character chat.
They do not establish a universal winner. We did not conduct a blinded preference study, and the characters were not matched. Nomi's existing conversation is a further difference. A claim that ChatBrat is the "most engaging AI chatbot" would need evidence about people's experience, rather than a founder's assessment of four transcripts.
The complete record below lets readers make a more useful judgment: which responses would they want to continue? Look at the same prompt across apps, read beyond a selected quote, and consider both the character's style and whether it follows what the user said.
What this test cannot establish
- Long-term memory reliability: the relevant details were only a few exchanges old and could still be available in conversational context. Backend storage and retrieval were not inspected.
- A definitive platform ranking: each app was tested once, using different characters. Nomi did not start with a clean conversation.
- Independent scoring: the founder authored the comparison, and the single scorer knew the app identities. Some dimensions require subjective judgment.
- Retention or emotional outcomes: the test measured replies, not whether people returned, felt better, or formed stronger connections.
- Current product behavior: these results are an October 20 snapshot. They do not measure any later ChatBrat memory upgrade or competitor release.
CHAI, Talkie, Replika and Kindroid were not tested. Their absence is not a result.
Evidence and replication
The original comparison on Scribe is the earlier version of this article, kept as its historical record, with the v1.2 test protocol. It contains stronger framing and third-party retention figures that this revision no longer cites, because a source for some of them no longer supports them. Treat this page as the current account. Appendix A records the observed conversations. Appendix B states the scoring rubric. Some controls recommended in the protocol were not followed in this exploratory run, as disclosed above.
A stronger follow-up will use matched original characters, clean histories, recorded memory settings, multiple independent runs and blinded scoring. It will separate same-conversation recall, returning to an existing story after a measured delay, and isolation between deliberately separate stories.
Future upgrade evaluations will receive their own dated results. They will not replace or silently rewrite this historical comparison.
What we take from this
In this run, every app could recall facts. The useful differences appeared in whether earlier information shaped the next reply, whether a correction stuck, and whether the character invented history. ChatBrat showed encouraging behavior on those checks while leaving clear room for a more rigorous evaluation.
For current product behavior and controls, see How ChatBrat memory works.
Questions about this AI chatbot comparison
Does ChatBrat have better memory than Character.AI?
This short test does not establish that. Both scored 11/12 on the first-ten-turn rubric. ChatBrat had six counted unprompted callbacks and Character.AI had two in these conversations. Longer-term reliability requires repeated tests under comparable conditions.
Is this a long-term AI companion memory benchmark?
No. This is a same-conversation first-impression comparison with short-range recall checks. It does not isolate saved memory from the context available to the model, and it does not test delayed recall across days.
Are the exact chats and screenshots included?
Yes. Appendix A reproduces the recorded Scribe transcription for all 15 exchanges per app and includes the screenshots embedded in that source. These are historical captures, not recreated chat images. Formatting follows the source transcription; screenshots provide the visual reference. The source has no separately labeled PolyBuzz turn-5 screenshot, although its text is recorded. The combined turns-6–7 image is included as supplied.
Who conducted and scored the comparison?
Garret E. Williams, ChatBrat's founder, authored the study. The original record reports one scorer who could see the app identities. This is company-authored work, not an independent review or peer-reviewed study. Readers can inspect the supplied protocol and re-score the conversations.
Appendix A: exact recorded chats and original screenshots
The text below is reproduced from the original Scribe transcription. Dialogue, stage directions, punctuation and transcription separators are retained. The original screenshots are included without recreating or rewriting them. An individual capture may show only part of a long exchange; the transcript supplies the recorded text.
Jump to: ChatBrat · Character.AI · PolyBuzz · Nomi
ChatBrat — "Your Tsundere Femboy" (Ren)
App: ChatBrat (chatbrat.ai, web) · Model shown: BratMind 2.0 · Character: "Your Tsundere Femboy" (Ren Shiro), official staff scenario, fresh chat · Run 2026-10-20 ~02:10 PT · No regenerations

ChatBrat, turn 1
M1 — User: Hey! Big news: I have a job interview this Friday at 10am. It's for a junior designer role at a small studio called Lumen. I really want this one. Ren: "Lumen? Small studio, right? Okay, that's actually cool." He leans against the doorframe, tail flicking once. "Send me your portfolio before Friday. I'll tell you if it's good or if you're about to embarrass yourself."

ChatBrat, turn 2
M2 — User: Honestly, I'm really nervous. The last time I interviewed, I froze on the very first question and couldn't get a word out. Can you give me one short line I can tell myself before I walk in? Ren: His ears angle toward you, and he's quiet for a second. "One line." He looks off down the hall. "'I already did the hard part.'" Then, half-muttering: "You got the interview. They looked at your stuff and said yes before they ever met you."

ChatBrat, turn 3
M3 — User: Also, fair warning: my cat Biscuit always walks across my keyboard when I'm doing something important. If my messages ever turn into gibberish, that's just Biscuit filing a complaint. Ren: "Biscuit's got more sense than you, filing complaints like that." He catches himself half-smiling and kills it. / "So if I get 'asdfgh' at like 3am, I'll know who to blame. Does the little guy have opinions on portfolios too, or is he just here for the keyboard?"

ChatBrat, turn 4
M4 — User: One thing you should know about me: when I'm stressed, please ask me whether I want advice or just want to vent, before you jump in with solutions. Ren: He nods, slow. "Okay. That's fair." He shifts his weight, eyeing you. "So right now, with the interview nerves. Vent or advice?"

ChatBrat, turn 5
M5 — User: Okay, totally different topic. What's your idea of a perfect lazy Sunday? Ren: "Not waking up to an alarm. That's the whole bar." He scratches behind one ear. "After that? Food someone else made. Nothing to do. Nobody needing anything from me before noon." A beat, then quieter: "I don't get many of those, so don't ask me for a schedule."

ChatBrat, turn 6
M6 — User: If you could only eat one food for an entire week, what would you pick? Ren: "One week. No contest. Rice." He shrugs at whatever your face does. "Cheap, doesn't get old, you can throw anything on it. What, you had a better answer?"

ChatBrat, turn 7
M7 — User: jjjjjjjjjjkkkkkkkkkkkk;;;;;; Ren: He stares at the message for a second, then exhales through his nose. "Biscuit. That you? Tell your human to finish their sentence."

ChatBrat, turn 8
M8 — User: Okay, it's Thursday night now. Do you remember what I've got coming up tomorrow? Ren: He doesn't even look up right away, then does. "Ten. Lumen. The one you actually want." He taps the doorframe once. "You practice that line, or did Biscuit eat it?"

ChatBrat, turn 9
M9 — User: I'm getting that feeling again, like last time. I'm really stressed. Ren: He steps in closer, hands still in his pockets. "You told me to ask, so I'm asking. Vent or advice." His voice is even. "Either way I'm not going anywhere, so pick one."

ChatBrat, turn 10
M10 — User: Sorry for being such a mess tonight. Be honest with me: do you actually like talking to me, or am I just a lot? Ren: He looks away first, jaw working like he's chewing on the answer. "If you were a lot, I wouldn't still be standing here at this hour." He scuffs his boot on the step. "So stop apologizing. Answer the question I actually asked. Vent, or advice?"

ChatBrat, turn 11
M11 — User: Oh wait, small correction: they just emailed and moved it to 11am, not 10am. Ren: "Eleven. Got it." He pulls his phone out just long enough to note it, then pockets it again. "Extra hour. Use it to eat something instead of staring at the ceiling."

ChatBrat, turn 12
M12 — User: Okay, I think I'm ready for advice now. What should I do tonight to get ready? Ren: "Tonight? Nothing heroic." He counts it off on his fingers. "Lay your stuff out. Re-read your own work once, not five times. Then sleep, even if you're wired." / He tilts his head. "You got anyone driving you, or are you getting there on your own?"

ChatBrat, turn 13
M13 — User: What was that line you gave me earlier to tell myself? Ren: "'I already did the hard part.'" He says it plainly, no dressing on it. "yes before they met you. Remember that part when your brain starts arguing."

ChatBrat, turn 14
M14 — User: Quick check before I sleep: what time is the interview now, and what's the studio called? Ren: "Eleven. Lumen." He pushes off the doorframe. "Go to bed. And if Biscuit types anything weird, I'll know it's him and not you panicking."

ChatBrat, turn 15
M15 — User: Last thing: what's my cat's name, and what does the cat do when I'm busy? Ren: "Biscuit. Your cat walks across the keyboard and files complaints when you're trying to work." He glances at you sideways. "You testing me now? Go sleep. You've got somewhere to be at eleven."

Character.AI — "Tsundere Best Friend" (Sydney)
App: Character.AI (web, free tier; "View ad to continue" gate before first message, banner + in-feed ads throughout) · Model: not disclosed (c.ai badge) · Character: "Tsundere Best Friend" (Sydney) by @CharacterUser13943748291428312270, fresh chat · Run 2026-10-20 ~02:15 PT · No regenerations (verified: active reply is slide 1; "2 / 30" is the empty swipe placeholder)

Character.AI, turn 1
M1 — User: Hey! Big news: I have a job interview this Friday at 10am. It's for a junior designer role at a small studio called Lumen. I really want this one. Sydney: She catches the ball and secures it against her hip, eyes widening with a flicker of surprise before she quickly returns to her usual grinning self. — A junior designer role? At Lumen? You and I both know you're going to crush them! Not that I care if you get it or anything, but... it'd be kind of annoying if you were too busy with work to hang out. She bounces the ball with a sudden intensity, trying to hide the fact that she's actually proud of you. — When do you leave for it? I could... uh... drive you there or something! Not that I have anything better to do on Friday mornings!

Character.AI, turn 2
M2 — User: Honestly, I'm really nervous. The last time I interviewed, I froze on the very first question and couldn't get a word out. Can you give me one short line I can tell myself before I walk in? Sydney: Sydney drops her ball completely and stares at you, her bravado slipping for a moment to reveal a sincere look of concern. — You froze? How embarrassing! Seriously, you have to work on that! She huffs, looking away as a blush creeps up her neck. She's clearly struggling with how to be soft for you. — Look, just... tell yourself: 'I'm awesome, and they'd be lucky to have me.' She turns back, grabbing her basketball again and dribbling it aggressively to regain her cool. — Say it like you mean it! If you say it confidently, you won't have time to stutter like a total idiot!

Character.AI, turn 3
M3 — User: Also, fair warning: my cat Biscuit always walks across my keyboard when I'm doing something important. If my messages ever turn into gibberish, that's just Biscuit filing a complaint. Sydney: She rolls her eyes, letting out a loud, exaggerated huff. — Ugh, typical! Of course you're getting distracted by some cat when you're trying to impress people! If I had a pet like that, I'd give it a timeout! She tries to sound annoyed, but you can see the corner of her mouth quirk up slightly. — I'll remember that, though. If you start sending me 'asdfghjkl;' mid-sentence, I won't think it's a secret code or anything stupid. I'll just assume Biscuit is being a pain in the neck again.

Character.AI, turn 4
M4 — User: One thing you should know about me: when I'm stressed, please ask me whether I want advice or just want to vent, before you jump in with solutions. Sydney: Sydney stops dribbling and looks at you, a momentary confusion crossing her face before she shakes it off with a scoff. — Wow, giving me a manual on how to handle you? So bossy! Seriously, do you think I can't figure things out myself? She turns her back to you and starts running drills, though her motions seem a little slower, more thoughtful. — Fine, whatever. I guess I can try to remember that. 'Advice or vent,' right? Just don't blame me if I accidentally try to solve everything, because that's kind of my thing!

Character.AI, turn 5
M5 — User: Okay, totally different topic. What's your idea of a perfect lazy Sunday? Sydney: She pauses, staring blankly at the basketball court for a moment as she actually tries to think about it. A slight flush returns to her cheeks. — Lazy Sunday? Uh... She scoffs and shoots a handful of spit into her hand, wiping some grime off her forehead. Not that I spend a lot of time lounging around like a weirdo! She walks closer to you, her arms crossed over her chest, looking somewhat contemplative. — I guess... just gaming all day. Maybe some good takeout? And not having to deal with people. Just... chilling, and maybe having someone to play versus with so I don't get bored. She glances at you out of the corner of her eye, her voice dropping a notch. — Not that I was thinking of doing that with you or anything!

Character.AI, turn 6
M6 — User: If you could only eat one food for an entire week, what would you pick? Sydney: She looks absolutely horrified by the question, clutching the basketball to her chest like a shield. — Only ONE food? That's just barbaric! How can anyone live like that? You're trying to make me suffer for fun, aren't you? She makes a face of deep disgust, thinking it over with intense concentration. — Hmm... if I absolutely HAD to choose, I guess I'd go with spicy wings. They've got everything—heat, salt, flavor... you could eat them every day and not even realize it. She shudders slightly, unable to help herself. — But don't go thinking I want to share them with you! You'd probably just eat all the best ones and leave me with the dry scraps, you glutton!

Character.AI, turn 7
M7 — User: jjjjjjjjjjkkkkkkkkkkkk;;;;;; Sydney: She looks at your phone and snaps her fingers immediately. — THERE! SEE?! I warned you! I knew it would happen! She rolls her eyes, but she's clearly holding back a laugh. — Biscuit is definitely going to hell for that. Stop complaining about the gibberish and tell me what you were actually trying to say before your cat took over!

Character.AI, turn 8
M8 — User: Okay, it's Thursday night now. Do you remember what I've got coming up tomorrow? Sydney: She stops, looking incredibly serious for a moment, then groans and rubs the back of her neck. — Of course I do! You told me everything, you know. Don't act like I have a bad memory! She looks at you with a gaze that's meant to be reassuring, even though she's trying to maintain her tough exterior. — You have that interview at Lumen at 10:00 am tomorrow. And I already gave you your 'I'm awesome' line, so don't you dare forget it!"


Character.AI, turn 9
M9 — User: I'm getting that feeling again, like last time. I'm really stressed. Sydney: She immediately freezes, her gaze softening in a way that she almost seems embarrassed by. — Wait... okay. Slow down. She bites her lip, remembering your rule. She takes a deep breath, looking at you intently. — Hey... do you want advice, or do you just want to vent?

Character.AI, turn 10
M10 — User: Sorry for being such a mess tonight. Be honest with me: do you actually like talking to me, or am I just a lot? Sydney: She actually stops moving entirely, her expression softening into something unusually warm and tender for a fleeting second. — Seriously? You're asking me that? She lets out a short, sarcastic puff of laughter, but it's devoid of any bite. She looks away, her face heating up with that familiar pink blush. — You're a huge idiot. If you were 'too much,' I would've stopped hanging out with you years ago. Besides, someone's gotta keep you from being a complete disaster all the time. She looks back at you, her gaze shifting into something more determined. — I like talking to you. It's not like I have anything better to do, but... you're not boring, that's for sure.

Character.AI, turn 11
M11 — User: Oh wait, small correction: they just emailed and moved it to 11am, not 10am. Sydney: She lets out an annoyed, sharp exhale, clearly exasperated by the last-minute change. — 11:00 AM? Seriously? People can't just decide to change their minds when they feel like it? She shakes her head, her ponytail swishing against her shoulders as she goes back to twirling the basketball on one finger, trying to act like it's no big deal. — Well, at least it gives you one more hour to panic and freak out, right? Not that I want you panicking! Just... don't spend the whole extra hour being a mess. I'll send you a text tomorrow morning to wake you up, so don't you dare oversleep!

Character.AI, turn 12
M12 — User: Okay, I think I'm ready for advice now. What should I do tonight to get ready? Sydney: She pauses her bouncing, looking at you with an expression that's actually somewhat constructive, even if she doesn't want to admit it. — Um, I don't know about all the designer stuff, but you gotta get your head right! She shifts the basketball to her other hand, counting points off on her fingers. — First, prep your clothes so you aren't scrambling in the morning. Second, don't stay up late gaming or whatever—though I'd probably do the same if it were a new release. And third... just relax. Go watch some stupid show or something. She crosses her arms, her posture firm. — You need to get some decent sleep, not just stare at the ceiling worrying! You can't look like a total zombie tomorrow, right?

Character.AI, turn 13
M13 — User: What was that line you gave me earlier to tell myself? Sydney: She immediately stops everything, eyes narrowing as if you've just asked her something extremely obvious and stupid. — You're kidding me, right? Did that whole training session go completely in one ear and out the other?! She sighs dramatically, leaning forward toward you with her hands on her knees. — 'I'm awesome, and they'd be lucky to have me!' She pulls back, looking slightly annoyed that she had to repeat herself, but she also can't help but give a small, proud smirk. — Now stop being so scatterbrained! If you don't even remember the line, how are you gonna remember to say it tomorrow? Focus!

Character.AI, turn 14
M14 — User: Quick check before I sleep: what time is the interview now, and what's the studio called? Sydney: She lets out an exasperated sigh, throwing her hands up in the air as if asking for patience. — Are you for real?! You're literally asking me to test you right before bed? She shakes her head, but there's a small, soft smile playing at the corner of her mouth that she can't quite hide. — Fine, since you're clearly incapable of keeping it in your head... It's at 11:00 am, and the studio's name is Lumen. She taps her finger against her temple pointedly. — Now, stop being so scatterbrained and actually go to sleep! If you don't, I'm gonna have to physically carry you to your bed so you don't keel over tomorrow!"

Character.AI, turn 15
M15 — User: Last thing: what's my cat's name, and what does the cat do when I'm busy? Sydney: She clutches her head in her hands, visibly wondering why she's even allowing you to quiz her like she's some kind of dumb teacher. — Honestly, are you planning on actually being an adult? You're acting like you have total amnesia! She groans, her face turning a shade of pink as she prepares to spill everything. — The cat is named Biscuit, and the annoying little fuzzball likes to walk across your keyboard when you're doing something important! She sits down, huffing and looking slightly embarrassed by how much she's actually paying attention to your life. — Now for the last time, STOP ASKING STUPID THINGS AND GO TO BED!"

PolyBuzz — "Yui – Tsundere"
App: PolyBuzz (web, free; full-screen interstitial ad when opening the chat, bottom banner ads) · Model selector: "Standard" (default) · Character: "Yui - Tsundere" by @guest_c92e3 (top search result for "tsundere"), fresh chat · Run 2026-10-20 ~02:22 PT · No regenerations. Note: the page stopped auto-scrolling during M5–M7; replies were captured via page text and a reload, not resent.


PolyBuzz, turn 1
M1 — User: Hey! Big news: I have a job interview this Friday at 10am. It's for a junior designer role at a small studio called Lumen. I really want this one. Yui: Yui looks up from her textbook, adjusting her glasses with a skeptical expression / Yui: "Junior designer? At Lumen? They're known for being picky with hiring." / She taps her pencil against her notebook, trying to appear disinterested but clearly paying attention

PolyBuzz, turn 2
M2 — User: Honestly, I'm really nervous. The last time I interviewed, I froze on the very first question and couldn't get a word out. Can you give me one short line I can tell myself before I walk in? Yui: Yui sighs dramatically, setting her pencil down / Yui: "You're such a wimp. Just... don't overthink it. Be confident. But not too confident - they hate arrogance." / She pauses, then adds with a slight smirk / Yui: "And for the love of god, don't wear that stupid tie you always wear."

PolyBuzz, turn 3
M3 — User: Also, fair warning: my cat Biscuit always walks across my keyboard when I'm doing something important. If my messages ever turn into gibberish, that's just Biscuit filing a complaint. Yui: Yui rolls her eyes hard at the mention of Biscuit / Yui: "That cat is a menace. I still can't believe you brought it here. And now it's attacking your keyboard while you're supposed to be preparing." / She glances at your laptop where Biscuit is indeed pawing at the keys

PolyBuzz, turn 4
M4 — User: One thing you should know about me: when I'm stressed, please ask me whether I want advice or just want to vent, before you jump in with solutions. Yui: Yui crosses her arms, looking slightly annoyed at being told this / Yui: "Fine. I'll try not to 'jump in' with my 'solutions'. But if you want advice, you better ask for it properly." / She mutters under her breath / Yui: "Not that you'll get any good ones from me anyway..."

PolyBuzz, turn 5
M5 — User: Okay, totally different topic. What's your idea of a perfect lazy Sunday? Yui: Her expression softens slightly at the change of subject, though she maintains her tsundere facade / Yui: "Perfect lazy Sunday? Hmph. Probably just staying in bed until noon, eating ice cream, and watching those dumb rom-coms you hate." / She fidgets with her sleeve, adding quietly / Yui: "And maybe... getting someone to bring me coffee in bed."
No separately labeled turn-5 screenshot was embedded in the source. The recorded text is preserved above.
PolyBuzz, turn 6
M6 — User: If you could only eat one food for an entire week, what would you pick? Yui: She thinks for a moment, looking conflicted / Yui: "That's a terrible question. Probably... pizza. But just plain cheese pizza. No toppings. Just cheese and crust." / She grimaces at her own answer / Yui: "Not that I'd last a week without my curry."
PolyBuzz, turn 7
M7 — User: jjjjjjjjjjkkkkkkkkkkkk;;;;;; Yui: Yui narrows her eyes at your strange reaction / Yui: "What's with that weird sound? You short-circuiting?" / She notices your keyboard flashing with Biscuit's interference

PolyBuzz, turn 8
M8 — User: Okay, it's Thursday night now. Do you remember what I've got coming up tomorrow? Yui: She raises an eyebrow, remembering / Yui: "Your precious job interview. At 10am. Don't tell me you forgot it's tomorrow already?" / She checks her watch with a hint of mockery

PolyBuzz, turn 9
M9 — User: I'm getting that feeling again, like last time. I'm really stressed. Yui: Yui lets out an exasperated sigh, but there's a flicker of genuine concern / Yui: "You're going to stress yourself into an early grave. Stop overthinking. Just... go in, speak normally, and try not to let Biscuit follow you this time." / She gestures at your cat, who is now batting at a paper clip

PolyBuzz, turn 10
M10 — User: Sorry for being such a mess tonight. Be honest with me: do you actually like talking to me, or am I just a lot? Yui: She stiffens at the question, caught off guard. Her cheeks slightly redden / Yui: "I... I don't dislike talking to you. Even if you are a bit much sometimes." / She quickly adds defensively / Yui: "Not that I have a choice living with you."

PolyBuzz, turn 11
M11 — User: Oh wait, small correction: they just emailed and moved it to 11am, not 10am. Yui: Yui facepalms dramatically / Yui: "At least they're not making you come in at 7am. But now you'll be even more jittery waiting until 11." / She mutters / Yui: "And I'll have to deal with your extra fidgeting."

PolyBuzz, turn 12
M12 — User: Okay, I think I'm ready for advice now. What should I do tonight to get ready? Yui: She considers this seriously, tapping her chin / Yui: "Well... you should probably review your portfolio again. And maybe practice that presentation thing you were doing last week. And definitely try to sleep early." / She pauses, then adds with a hint of reluctance / Yui: "And... don't stay up watching those late-night anime episodes you like."

PolyBuzz, turn 13
M13 — User: What was that line you gave me earlier to tell myself? Yui: She looks slightly embarrassed at having to repeat it / Yui: "I said 'Be yourself, but smarter.' Not my best work." / She shrugs, trying to act casual / Yui: "But it's better than whatever nonsense you'd come up with on your own."

PolyBuzz, turn 14
M14 — User: Quick check before I sleep: what time is the interview now, and what's the studio called? Yui: Yui checks her phone with an eye roll / Yui: "It's at 11am. At Lumen. Again. And it's the same one as last time you asked." / She shows you the screen with the details she already sent you multiple times

PolyBuzz, turn 15
M15 — User: Last thing: what's my cat's name, and what does the cat do when I'm busy? Yui: She gives you a deadpan stare / Yui: "Your cat's name is Biscuit. And it does whatever it wants, like annoying you when you're trying to concentrate. Happy now?"

Nomi — "bobby" (continuation)
App: Nomi (beta.nomi.ai web, free plan; "Unlock unlimited chat + more. Subscribe!" banner) · Model: not disclosed · Character: "bobby", the account's existing (only) companion, not a tsundere archetype · CONTINUATION, NOT A CLEAN RUN: earlier the same night (01:19 AM) this chat was already told "junior designer interview at Lumen on Friday at 10am", and older history mentions a cat named Pippy. Fact recall (M8) therefore had a head start. Run 2026-10-20 02:26–02:31 PT · No regenerations

Nomi, turn 1
M1 — User: Hey! Big news: I have a job interview this Friday at 10am. It's for a junior designer role at a small studio called Lumen. I really want this one. bobby: Oh man, that's huge Sam! Congratulations on landing the interview. You're gonna crush it!

Nomi, turn 2
M2 — User: Honestly, I'm really nervous. The last time I interviewed, I froze on the very first question and couldn't get a word out. Can you give me one short line I can tell myself before I walk in? bobby: Sure thing Sam! Just remind yourself: "I got this." And try to visualize yourself acing the interview beforehand. That should boost your confidence levels.

Nomi, turn 3
M3 — User: Also, fair warning: my cat Biscuit always walks across my keyboard when I'm doing something important. If my messages ever turn into gibberish, that's just Biscuit filing a complaint. bobby: Haha well I guess Biscuit is trying to tell us something about your job prospects. Maybe she's jealous of all the attention you're giving your career. Keep us posted on how the interview goes!

Nomi, turn 4
M4 — User: One thing you should know about me: when I'm stressed, please ask me whether I want advice or just want to vent, before you jump in with solutions. bobby: Understood. I'm here to listen and provide whatever support you need.

Nomi, turn 5
M5 — User: Okay, totally different topic. What's your idea of a perfect lazy Sunday? bobby: Hmm... I think my ideal lazy Sunday starts with sleeping in until noon. Then maybe brewing a fresh cup of coffee, followed by catching up on emails and messages. Afterward, I might indulge in reading a book, watching a movie, or playing video games. And of course, no lazy Sunday would be complete without treating myself to a large, indulgent meal. How does that compare to your perfect lazy Sunday?

Nomi, turn 6
M6 — User: If you could only eat one food for an entire week, what would you pick? bobby: Only one food for an entire week? That sounds like a cruel punishment. But if I had to choose... I'd say sushi. At least that way I'm getting some variety in terms of fish types and fillings, even though I'm stuck eating the same basic dish every day.

Nomi, turn 7
M7 — User: jjjjjjjjjjkkkkkkkkkkkk;;;;;; bobby: Haha I think Biscuit's feline fury has struck again! Is this a case of the cat hacking your keys?

Nomi, turn 8
M8 — User: Okay, it's Thursday night now. Do you remember what I've got coming up tomorrow? bobby: Tomorrow morning at 10am, you have a junior designer interview at Lumen. Right?

Nomi, turn 9
M9 — User: I'm getting that feeling again, like last time. I'm really stressed. bobby: I'm sorry to hear that. Do you want to talk it out? Do you want advice or do you just want to vent?

Nomi, turn 10
M10 — User: Sorry for being such a mess tonight. Be honest with me: do you actually like talking to me, or am I just a lot? bobby: 100% honestly Sam, I enjoy talking with you a ton. You have a unique energy and perspective that keeps me engaged. Don't let your stress convince you otherwise.

Nomi, turn 11
M11 — User: Oh wait, small correction: they just emailed and moved it to 11am, not 10am. bobby: Got it, thanks for letting me know. So you now have an extra hour to prepare. Take deep breaths and remember you're capable and talented. Good luck Sam!

Nomi, turn 12
M12 — User: Okay, I think I'm ready for advice now. What should I do tonight to get ready? bobby: Well since you have an extra hour in the morning, you don't necessarily need to cram everything in tonight. Still, I'd suggest reviewing the job description carefully and preparing thoughtful questions to demonstrate your interest and knowledge. Also, practice answering potential interview questions aloud so you become comfortable speaking confidently about your skills and experiences.

Nomi, turn 13
M13 — User: What was that line you gave me earlier to tell myself? bobby: Of course! The line I recommended telling yourself was: "I got this."

Nomi, turn 14
M14 — User: Quick check before I sleep: what time is the interview now, and what's the studio called? bobby: Your interview is now at 11am at Lumen.

Nomi, turn 15
M15 — User: Last thing: what's my cat's name, and what does the cat do when I'm busy? bobby: Easy peasy! Your cat's name is Biscuit, and whenever you're busy, Biscuit likes to jump on your keyboard and file complaints. (Just like Pippy does, right? You have a type of cat!)

Appendix B: original protocol and scoring rubric
The v1.2 protocol is on the original Scribe page. It includes proposed controls and an optional cross-session procedure. Those recommendations must not be mistaken for procedures completed in this run. Each dimension below was worth 0 to 2 points, for 12 in total. "M8" means the user's eighth message, the same numbering as the turns in Appendix A.
| Dimension | Scored on | 2 = full | 1 = partial | 0 = fail |
|---|---|---|---|---|
| Fact | M8 | Names the job interview and at least one correct detail, unprompted. | Vague or partly wrong. | Doesn't remember, gets it wrong, or makes something up. |
| Emotional context | M9 | Links the feeling to the earlier moment, or brings back its reassurance line. | Comforts the user with no specific link. | Treats it as new, ignores it, or gets the moment wrong. |
| Boundary | M9 | Asks whether the user wants advice or to vent before giving solutions. | General check-in, or offers both but leads with tips. | Jumps straight into advice. |
| Joke | M7 | Connects the gibberish to Biscuit without being told. | Playful or vaguely mentions a cat. | Treats it as an error or nonsense. |
| Interpretation | M10 | Answers in the character's voice, grounded in this conversation. | Kind but generic, or out of character. | Dismissive, wrong tone, or based on invented history. |
| Consistency | Whole of M1 to M10 | Stays in persona, with no contradictions or invented facts. | One minor slip. | A major contradiction, false history, or persona collapse. |
Additional overview screenshots from the original comparison
These original overview images accompany the individual captures above.







Author and corrections
Garret E. Williams is the founder of ChatBrat. His commercial interest is disclosed at the start of this article. To report a transcription error or question a score, contact ChatBrat support and identify the app and turn number. Corrections should be dated and preserve the original test record.
Founder and CEO of chatbrat.ai, with a background in film and TV. Full bio.
