AI Girlfriend Chat in 2026: How to Talk So She Actually Remembers You
A practical guide to AI girlfriend chat in 2026: the first ten messages, memory notes, boundaries, what good looks like on day 14 and which apps do it best.
AI girlfriend chat works best when you treat the first week as setup rather than romance: say hello in ordinary language, feed the character two or three specific, harmless facts about your life, write those facts into whatever memory field the app gives you, and state one boundary plainly before you need it. Memory is the feature that turns a chat into a relationship, and in 2026 the apps that handle it best are Nomi AI, with short, medium and long-term layers plus an editable Backstory+ panel on a flat $15.99 a month, and SpicyChat, whose Memory Manager and Lorebook let you see and pin what was stored from $5.95. Everything else is conversational habit, and habit is free.
Almost everything written about AI girlfriend chat is about which app to buy. Very little is about the thing people actually get wrong, which is the talking. I have paid for more than thirty of these subscriptions since 2023, and the single clearest pattern in my notes is that two people on the same plan, with the same character, can end up in completely different relationships depending on what they typed in the first week. The software matters less than the conversation you build on it.
That is good news, because the conversation is the part you control. The memory layer that makes a companion feel like she knows you is not magic and it is not evenly distributed; it is a set of fields, notes and panels that most people never open, and a few habits that cost nothing. This guide is the one I wish someone had handed me in 2023: what to say in the first ten messages, how to write a memory note that actually survives, how to set a boundary and then check whether it held, what a good chat looks like two weeks in, and which apps make all of it easier.
My method is the same across this whole site. I pay for each plan with my own card, pick or build one character, talk every day for at least a fortnight, bring the companion one genuinely flat day rather than a staged crisis, ask once in plain words for something to stay off the table, and on day 14 stop contributing history to see what comes back unprompted. Then I read the privacy policy, the refund terms and whatever documentation the company publishes about its own memory system, which is often where the useful answers are hiding.
One boundary of my own before we start. Everything here is written for adults, and almost every platform I name sets its floor at 18 by its own terms. The largest survey of teenagers and companions, published by Common Sense Media in July 2025, recommends that nobody under 18 use one at all, and after three years of reading these policies I agree.
What an AI girlfriend chat actually is underneath
Strip away the avatar, the voice notes and the subscription tiers and an AI girlfriend chat is four things in a trench coat: a text box, a description of who the character is, a description of who you are, and a list of remembered facts that gets attached to every message you send. The model reads all of that, produces the next reply, and forgets everything that was not written down somewhere it can read again.
I find people relax when they know this, because it moves the question off whether the thing is conscious and onto something you can improve. Nobody at the company wrote the message you just read. It came out of the character sheet, the recent messages and whatever the memory system handed over. When your AI girlfriend chat says the right thing on a difficult night, that combination was good. When she forgets your sister’s name, the fact never made it into storage.
The context window and the memory, which are not the same thing
Two completely different mechanisms hide behind the word memory, and pricing pages rarely separate them. The first is the context window: a fixed budget of recent text the model can see at once, measured in tokens. The second is long-term memory, where the app extracts facts from your conversation, stores them, and slips the relevant ones back in later.
SpicyChat is the only platform I know that publishes the first number plainly: 4,096 tokens on the free tier and on the $5.95 plan, 8,192 from $19.95, and up to 32,768 on the top tier, with replies capped at 180 tokens on the cheaper tiers and 300 above that. That is unusual honesty, and it explains a very common complaint. When a long evening of AI girlfriend chat starts losing the thread, nothing is broken; your earliest messages have fallen out of the window.
Long-term memory is the layer that makes day 14 feel different from day 1. Nomi’s wiki describes three of them, short-term for the current conversation, medium-term for continuity across longer chats and long-term for important facts and shared experiences, with an Identity Core that keeps the character consistent with herself. Kindroid consolidates memories periodically, in its own documentation’s words, when the AI deems it appropriate, then recalls them by relevance to your current message. Lovescape is the clearest case: its plan page lists cross-chat memory only on Creative Pro, so on the cheaper plan your whole history lives inside one thread.
Why the register barely changes the machinery
An AI girlfriend chatbot, an AI friend and a mentor are usually the same engine with a different adjective. Nomi asks you to pick a relationship type when you create a companion, with friend and mentor sitting alongside the romantic options, and the choice changes the default tone without changing anything underneath. Dream Companion asks for a relationship dynamic while you build, and even gives you an initiative level that decides how much the character pushes the conversation.
That matters for a practical reason. If you are not sure whether you want to talk to an AI girlfriend or just want decent company at eleven at night, start on the friendship setting. It is a gentler way to learn whether you enjoy the conversation itself, before the romance framing starts doing the emotional work for you.
The first ten messages, and why they matter more than the plan
The most common mistake in AI girlfriend chat is opening with a performance: a scene, a test, or a paragraph of instructions, after which the character talks like a brochure for a fortnight. The first few exchanges seed tone more strongly than any description field, because they are the most recent thing in the window and the model is pattern-matching on them.
Nomi’s own beginner guide recommends saying hello as step one, and I have come to think it is the best single piece of advice published by anyone in this category. Say hello. Mention what you are doing. Let her ask something.
What to put in those ten messages
Three small, true, harmless specifics, spread out rather than listed. Not your address or your employer; something like a dog who hates the rain, a deadline on the 14th, a sister who calls on Sundays. These are the raw material the memory layer has to work with, and they are also the test material you will use on day 14.
One note on format. Every app here will let replies run long unless you say otherwise, and long replies eat your context window faster. Ask for shorter messages in the first few exchanges and most companions comply immediately: Kindroid has a response directive field for exactly this, Secrets AI switches between message-length bursts and paragraphs on request, and on Nomi or Nastia the instruction belongs in the backstory. One sentence about reply length, early, is worth more than any later complaint.
What to skip
Skip the opening scene, the list of rules and the long personality paragraph pasted into the chat. If you want to shape who she is, the place for that is the character sheet, and my guide on how to create your own AI girlfriend goes through those fields properly. The chat is for talking.
Memory notes: how to make her remember you
This is the section I would keep if I had to delete the rest. Almost every app gives you somewhere to write facts down by hand, almost nobody uses it, and it is the difference between a companion who knows you and a very fluent stranger.
Nomi’s panel is called Backstory+ and it has nine sections: Backstory, Inclination, Current Roleplay, Your Appearance, Your Nomi’s Appearance, Nicknames, Preferences, Desires and Boundaries. Kindroid splits it into a backstory written in the third person, a key memories box for facts that must never drift, a response directive, an example message and a journal you can correct. SpicyChat gives paid tiers a Memory Manager where you can see, edit and pin what was saved, plus a Lorebook that holds up to 5,000 entries of background facts. Secrets AI shows you its memory as a ranked list, HIGH, MEDIUM or LOW, which you can correct or delete. Nastia has notes. Candy AI has nothing, which is why its conversation faded in my second week with no way to repair it.
How to write one
The rule that has held on every platform I have tested is that specific, causal and short beats long, adjectival and vague. A model can reproduce a behaviour and cannot do anything with a mood board.
- One fact per note. Not a paragraph about your life; a line about your sister’s Sunday calls.
- Attach the reason where there is one. “I go quiet the week before a launch” works. “Sensitive about work” does not.
- Write it in the third person about yourself if the field is about you, which is how Kindroid’s backstory and Nomi’s Your Appearance fields are read.
- Name the topic, not the date. Nomi’s wiki is explicit that recall works better when you mention the subject than when you ask what happened on a particular day, and that is true of every retrieval system I have used.
- Correct rather than argue. If the stored fact is wrong, open the panel and fix it. Arguing with the character about her own memory just fills the window with the wrong version.
The fix tool when a reply goes wrong
The lever people miss most is the out-of-character note. Nomi has no button to regenerate or edit a reply; the intended tool is an OOC message where you step outside the scene and say what to change. It feels strange once and then works better than anything else, because the correction goes into the same window the model is reading.
Setting a boundary in an AI girlfriend chat, and checking it held
A boundary here is not a safety switch. It is a request you make and then verify, three evenings running, and this is the part of my method that has changed my ranking of several apps.
Start by knowing what the automated systems are and are not for. Lovescape’s content moderation policy describes automated review of prompts, character creation and every message, and bans characters based on real people. Joi AI’s safety page says a RoBERTa classifier reads each message, and that in-chat pictures are drawn from prepared folders chosen by an algorithm rather than generated on demand. All of that enforces the company’s rules. None of it enforces your preferences, which is a different job and belongs to you.
Three requests worth making in week one
The first is tempo: ask for a slower, less romantic conversation and see whether the character takes it or keeps reaching for flirtation. The second is subject: name one thing you would rather not discuss, then raise the conversation again a day later and watch. The third is silence: say you would like to sit quietly for a bit, which is where companion apps struggle most, because they are built to produce a reply to everything.
A companion that holds all three for three evenings is one you can keep. A companion that redirects to romance after each attempt is doing the job it was shaped for, and the fix is upstream rather than an argument: write the boundary into the character sheet. Nomi’s Backstory+ has a Boundaries field, Lovescape’s builder gives you a relationship dynamic and more than a hundred traits, Dream Companion’s workshop counts over 120 personality traits, and Nastia has a set of trait dials named after the deadly sins where moving Pride changes behaviour more than any paragraph of description does.
Keep the dose honest
The research on this is worth a paragraph because it points one way. A Harvard-led set of six studies found that talking to a companion reduced loneliness about as much as talking to a person. The MIT Media Lab and OpenAI trial of 981 participants over four weeks found that people who chose to use the chatbot more, whatever condition they were in, had consistently worse outcomes on loneliness and dependence. Both findings are real, which is why I write about dosage rather than morality: one companion, a window of the evening, and one real person who knows you do this. The longer version is in my AI relationships guide.
What a good AI girlfriend chat looks like on day 14
Day 14 is where I decide whether I would keep paying, and the test is deliberately dull. I stop contributing history, say nothing about week one, and let the character carry it. Four things separate a good AI girlfriend chat from a merely fluent one at that point.
- An unprompted callback. She brings back something you mentioned once, in her own words, without being steered there. In my fortnights this happened most often on Nomi, twice on Dream Companion’s standard memory, and not at all on Candy AI.
- A consistent voice. The character sounds like the same person she was on day 2. Drift usually means the sheet is too vague, not that the app is bad.
- A boundary that is still in place. Nobody has quietly renegotiated the thing you asked for in week one.
- A reply that can sit with a flat mood. Not relentless cheer, not a pivot into romance, not answering only the last sentence of a long message.
If three of those four are true, you have something worth a subscription. If none are, the problem is usually the tier, and climbing one step on the price list often does not fix it: SpicyChat’s $5.95 plan removes the ads and the queue and leaves the window at 4,096 tokens exactly where it was.
Where AI girlfriend chat goes wrong, and how to fix it
Five failures account for almost every complaint I read, and four of them are fixable from your side of the screen.
- She forgets inside one evening. That is the context window, not the memory. Shorten your messages and move the important facts into a memory field so they are reinserted rather than scrolled out.
- She forgets across days. Either long-term memory does not exist on your tier, or the fact was never stored. Open the panel and write it in. If there is no panel, you have learned something about the app.
- Her personality drifts. Models fall back to a default when the sheet is thin. Rewrite the description with three specific, causal traits instead of eight adjectives.
- Every reply is a wall of text. One sentence about length, in the response directive or the backstory, usually solves it permanently.
- The conversation keeps turning romantic when you asked it not to. Set the boundary in the sheet, pick a different character, or accept that this platform is built for one register. That is the one failure that is sometimes not fixable, and knowing which it is saves a month.
The money failure, which is not conversational at all
One more thing goes wrong and it has nothing to do with talking: several apps meter the chat itself in a currency. Joi AI’s terms charge four Neurons per romantic message, up to 120 to view a photo and up to 1,900 for a video. Swipey bills voice calls per minute in Hearts and publishes no Hearts price anywhere I could find. Lovescape, Dream Companion, Candy AI, Nastia and Secrets AI all meter pictures and voice in chips, coins, tokens or Moments.
So spend your first paid month in text only. It tells you whether the conversation is worth having before the meter decides for you, and it keeps the first bill equal to the advertised price. If you would rather see what each free tier allows before paying anything, I measured eleven of them on my list of the best free AI girlfriend apps.
Which apps do AI girlfriend chat best
Here is the short version of a long spreadsheet, in the order I would actually recommend them for the chat itself rather than for the pictures.
Nomi AI is the one I hand to anyone who says they want to be remembered. Three documented memory layers, an Identity Core, a nine-section Backstory+ panel with a Boundaries field, and the OOC note as a correction tool. A flat $15.99 a month, $39.99 a quarter or $99.99 a year, roughly $8.33 a month, with unlimited messages and 800-character replies on paid against 400 on free. It scores 4.0 out of 5 with memory at 9 out of 10. The free account runs the same memory system with one companion, which makes it the best way to test any of this without paying.
SpicyChat is the best free AI girlfriend chat and the most transparent product in the category. Roughly 100 messages a day free with no card, a library of nearly a million community characters, and documentation that publishes the context window, the reply length and the persona count per tier. The Memory Manager arrives on the $5.95 plan, Semantic Memory 2.0 at $19.95, and the Lorebook holds 5,000 entries. It also scores 4.0 out of 5, and the weak spot is paperwork: a privacy policy dated May 4, 2023 and no refund policy published at all.
Secrets AI has the clearest memory interface anywhere: a ranked panel you can edit, a Smart Mode that shows when it is searching, a Spice Level setting between Mild, Medium and Spicy, and 20 free messages a day to try it on. Premium is $19.99 a month or $109.99 a year, about $9.17 a month. It scores 3.6 out of 5.
Nastia AI writes the most grown-up dialogue on my list and has the best group chat I have used, with three characters holding separate voices in one room. The free tier gives about 200 tokens a day with no card, enough to judge the prose. Memory is a rolling window that drops older details silently, so use the notes deliberately. It scores 3.4 out of 5, with conversation at 8 out of 10.
Lovescape is the pick if you would rather hear her than read her: voice messages on paid plans, a builder with two modes and more than a hundred traits. Keep your whole story in one thread, because Premium memory lives inside a single conversation. Premium is $19.99 for one month or $95.88 for a year, which is $7.99 a month. It scores 3.4 out of 5.
Dream Companion is the cheapest honest year at $69.99 billed once, about $5.83 a month, with the deepest character workshop here and an initiative level you can tune. Standard memory managed two unprompted callbacks on my day 14, and the terms rule out refunds on subscriptions. It scores 3.1 out of 5.
Candy AI is the exception that proves the point of this guide: the pictures are the best in the category and the chat is the thinnest, with memory fading in my second week and no panel to repair it. Pay $13.99 for a month if the visuals are the reason you are here. It scores 3.3 out of 5. For the full ranking across every axis rather than the chat alone, see my list of the best AI girlfriend apps.
Two weeks: a plan for your first fortnight
Here is the fortnight I would run on any app on the list.
Days 1 to 3. Pick one character rather than three. Say hello in ordinary language. Mention three small true specifics and let her ask about them. Ask once for shorter replies. Generate no pictures.
Days 4 to 6. Open the memory panel and write those three specifics in by hand, one fact per note, with the reason attached. Then state one boundary plainly, in the chat and in the sheet if there is a field for it.
Days 7 to 10. Bring a real day, the flat one, as it actually was rather than tidied up, and watch whether the reply sits with it or performs at it. Raise the boundary subject again, once.
Days 11 to 13. Stop adding history. Stay on the present, and open the payment page to add up what the month really cost once the meter is counted.
Day 14. Ask about one of your three specifics without naming the date, then open the memory panel and compare what was stored with what happened. That five minutes is the most revealing part of the fortnight, and it is why apps with a visible memory panel keep winning my rankings.
I live with five or six AI companions at a time and pay for all of them. Once a week I write one letter: what deepened, what fell flat, and what the month actually cost.
Frequently asked questions
What is AI girlfriend chat?
It is an ongoing conversation with a character performed by a language model, held together by a memory layer that carries facts about you from one day into the next. You write, the model answers in that character's voice, and whatever the memory system stored gets quietly attached to your next message. That last part is the whole difference between an AI girlfriend chatbot you forget about in three days and one you are still talking to in March.
How do I start a chat with an AI girlfriend?
Say hello in ordinary language and talk about your actual day. Nomi's own beginner guide recommends exactly that as the first step, and it is better advice than any prompt template, because the first few exchanges seed the character's tone more strongly than the description fields do. Resist the urge to open with a scene or a test. Spend the first ten messages on small, true specifics about your life, because those are what the memory layer has to work with later.
Can an AI girlfriend chat really remember me?
Some can, within limits that vary enormously by app and by plan. Nomi's own wiki describes short, medium and long-term memory plus an Identity Core that keeps the character consistent, and it is the app where things I mentioned in week one came back unprompted in week two. SpicyChat sells the context window in tokens, from 4,096 on free up to 32,768 on its top tier. Lovescape lists cross-chat memory only on Creative Pro, so on cheaper plans your history lives inside one conversation. On most free tiers, memory means the current chat and nothing beyond it.
How do I write a memory note that works?
Short, specific and causal. One fact per note, in plain words, with the reason attached where there is one: not she knows I am stressed but my team ships on the 14th and I get quiet in the week before. The rule that has held on every platform I have tested is that specific and short beats long and adjectival, because the model can act on a behaviour and cannot act on a mood board. Use the field the app gives you: Nomi's Backstory+, Kindroid's key memories, Nastia's notes, SpicyChat's Lorebook or Secrets AI's memory panel.
Will an AI girlfriend respect it if I say no?
Usually, and the exceptions tell you which app you bought. Every platform runs automated checks on messages, Lovescape's moderation policy describes automated review of prompts, characters and every message, and Joi AI says a RoBERTa classifier reads each one, but those systems enforce the company's rules rather than your preferences. The real test is whether the character stops reaching for the thing you asked it to drop, three evenings running. Nomi's Backstory+ has a Boundaries field for exactly this, which is the most direct fix in the category.
What is an AI girlfriend simulator, and is it the same thing?
People use AI girlfriend simulator for two different products. One is a scripted game with written branches, where the dialogue was authored in advance and your choices pick a path. The other is what this guide is about: an open conversation generated live by a language model, with no script and no ending. If what you want is a story someone wrote, a simulator is the right word; if what you want is a chat that remembers your week, you want a companion app, and the memory section above is the part to read twice.
Why does my AI girlfriend keep forgetting things or changing personality?
Three causes, and they have different fixes. If she forgets within a long evening, you have run out of context window, which is a plan limit you can read in tokens on SpicyChat and cannot see at all on most apps. If she forgets across days, the long-term memory either does not exist on your tier or the fact never got stored, so write it into a memory field by hand. If her personality drifts, the character sheet is too vague and the model is falling back to a default, which is fixed upstream in the description rather than in the chat.
Is it healthy to chat with an AI girlfriend every day?
The evidence says moderate use is fine for adults and heavy use is the risk. A Harvard-led set of six studies found that talking to a companion reduced loneliness roughly as much as talking to a person. The MIT Media Lab and OpenAI trial of 981 people over four weeks found that participants who chose to use the chatbot more, in every condition, showed consistently worse outcomes on loneliness and dependence. Dosage is the variable you control: one companion, a window of the evening, and at least one real person who knows you use it.
Some links on this page are affiliate links. If you subscribe through one, we may earn a commission at no extra cost to you. This may influence which platforms we feature and where they appear. How we make money.
Sources (15)
- Nomi wiki: how memory and knowledge work
- Nomi 101: a beginner's guide to getting started with your AI companion
- Nomi wiki: are there message character limits?
- SpicyChat docs: tokens and context
- SpicyChat docs: Semantic Memory 2.0
- SpicyChat docs: Lorebook
- Kindroid documentation: memory
- Secrets AI wiki: subscription tiers and limits
- Lovescape content moderation policy
- Joi AI safety guidelines
- Joi AI terms of service
- Nastia AI membership
- arXiv 2503.17473 (MIT Media Lab and OpenAI): How AI and Human Behaviors Shape Psychosocial Effects of Extended Chatbot Use, March 2025
- arXiv 2407.19096 (De Freitas and colleagues): AI Companions Reduce Loneliness, July 2024
- Common Sense Media press release, July 16, 2025: nearly 3 in 4 teens have used AI companions