Custom AI Girlfriend: Which Settings Actually Change Her
A custom AI girlfriend responds to two controls that genuinely change her behavior, and to several that mostly do not. The persona text and the system prompt rewrite what the model is asked to produce. Trait sliders compress into a few words of that same text, appearance touches nothing in the conversation, and the memory switch outweighs all of them.
Key Takeaways:
- The memory switch changes more than the rest of the settings page combined, because it decides whether anything you said yesterday exists today
- A persona description and a system prompt are both text placed in front of the conversation, which makes them the only controls that speak to the model in its own currency
- Trait sliders have to become words before a model can read them, and neighboring values on a 10-point scale often become the same words
- Appearance settings feed an image system that never sees your messages, so they cannot change a single reply
- A companion assembles text to match its instructions. Turning every dial toward warmth produces a warmer file, not a warmer someone
What Makes a Custom AI Girlfriend Custom
Written instructions move her, and everything else is a wrapper around written instructions. Whatever a settings page looks like, the model only ever receives one thing: a block of text. The persona description you wrote and the system prompt the app supplies both go into that block ahead of the conversation, which is why they carry weight no slider can match.
The system prompt is the app’s half. It sets the rules that apply to every character on the platform: stay in character, refuse certain topics, keep replies under a length, never claim to be human. Some apps expose part of it as a “custom instructions” field. Most do not, and when a company edits that hidden half on a Tuesday, thousands of characters change manner at once without a single user having touched anything.
The persona description is your half, and it is doing more than people expect from a text box. It arrives with every request, not once at creation, so a sentence like “she answers in 2 or 3 sentences and asks a question back about every third reply” is being applied to the message she is writing right now.
Both halves compete with the rest of the block. The architecture behind these models weighs every part of the input against every other part when producing each piece of the reply (Vaswani et al., 2017), and your 60-word description is sitting next to 40 messages of conversation you wrote in your own voice. Late in a long chat, your accumulated tone outweighs the instruction. The character drifts toward you, and it looks like a mood.
When she goes bland by message 60, resist rewriting the description. Open a fresh chat with the identical text and see whether she comes back sharp. If she does, dilution was the problem.
What Each Setting Plausibly Does
The settings page presents controls as equals. Sorted by what they can actually reach, the list is lopsided. That lopsidedness is the whole case for creating your own ai girlfriend personality-first.
| Control | What it most likely does underneath | Effect on her replies | How to check it |
|---|---|---|---|
| Persona description | Text inserted before every conversation | Large, and it is the one thing you fully author | Rewrite one line, open a fresh chat, compare |
| System prompt or custom instructions | Text inserted before your description | Large, and it can override you | Ask for something the app forbids and read the refusal wording |
| Memory on or off | Facts retrieved from a database and pasted in | Largest of all, but only after there is something stored | Mention a detail today, ask about it in a new chat tomorrow |
| Trait sliders | Values converted into adjectives or preset sentences | Small, and often nothing between neighboring values | Same 3 prompts at the minimum and at the maximum |
| Relationship or scenario preset | A swap of preset paragraphs | Moderate, because it replaces text wholesale | Read what changes in her opening message |
| Appearance and avatar | Parameters for a separate image system | None on the text | Change her hair, send the same message, compare |
| Voice | Speech synthesis applied after the text exists | None on wording, large on how it lands | Read the transcript with the audio off |
Two rows deserve a note. Presets often do more than sliders despite looking flimsier, because a preset swaps in a whole paragraph of instruction rather than adjusting a number. And voice sits at zero for content while changing the experience considerably, which is the one place where a cosmetic control is worth real money to the right person.
Read that table as a spending guide for your attention. Ten minutes in the description field is worth more than an hour anywhere else on the page. That is where how to create an ai girlfriend says to spend it too.
Why the Sliders Compress
A slider has to become language before the model can act on it, and language does not have 10 distinct settings for “playful.” Somewhere in the app, a 6 and a 7 are being converted into text, and there are only so many ways to write that sentence. Plenty of implementations map several adjacent values onto identical wording.
Three implementations are plausible from the outside, and they behave differently:
- Adjective substitution. The value picks a word, so “playfulness: 8” becomes “very playful” in the description. Coarse, and the gaps between steps are mostly invisible.
- Preset sentences. The value picks one of 3 or 4 written instructions, each specifying behavior. Better, because a sentence can say what the character does rather than what she is.
- Sampling settings. A control labeled something like “creativity” may change how the model picks among likely continuations. Real, measurable, and not a personality change: it makes replies more or less predictable, not warmer or colder.
None of this makes sliders a con. An adjective in the right place does shift output, and a preset sentence shifts it a lot. What sliders cannot deliver is the resolution their design promises, because a 10-point scale implies 10 states and the text underneath rarely has more than 3.
Sliders labeled with something like “honest” or “challenging” have an extra problem. They push against a tendency that was tuned into the model during training, when human raters preferred agreeable answers (Sharma et al., 2023), so the label changes the register far more reliably than it changes the verdict.
Set a slider to its extreme rather than to 7 out of 10. Extremes are the only values guaranteed to map to different text.
Why Memory Outranks Everything on the Page
Because it is the only control that changes the input rather than the instruction. Every other setting adjusts how the model is told to sound. Memory adjusts what it has to talk about, and material beats manner in any conversation that has run for more than a week.
The difference shows up in a specific kind of reply. With memory off, a warm answer has to be built entirely out of the message you just sent, so it comes back as a paraphrase of your own words in a nicer register. With memory on, the reply can contain something you did not just supply. That is the whole gap between a character who sounds interested and one who appears to have been paying attention. That gap is also the layer that decides what an ai girlfriend actually is rather than a chatbot with a name.
Which is why the timing of your verdict matters. An empty memory store on day 1 performs identically to memory being switched off, so the first evening tells you nothing about whether the feature works.
Testing a Setting Without Fooling Yourself
Changing one thing and reading one reply proves nothing, because you already know what you changed. Expectation does the rest. A method that survives that takes about 10 minutes.
Write down 3 fixed prompts before you touch anything: one factual, one about a small annoyance, one asking for an opinion you can check. Open a fresh chat, set the control to its minimum, send all 3, save the replies. New chat, same 3 prompts, control at maximum. Compare the pairs side by side rather than reading them in sequence, and change exactly one control between runs.
What does not work is asking her whether the setting changed anything. She has no view of her own configuration and will produce a confident, plausible account of a change that may not have happened. Treat that answer as writing, not as reporting. It is the same reflex behind an ai inventing a shared memory.
FAQ
Which setting actually changes how my AI girlfriend talks? The written description, and after that the memory switch. Both reach the model directly as text, while a slider only reaches it after being converted into a few words. If you can only edit one field, edit the one where you write full sentences about what she does.
Do personality sliders do anything at all? Some, at the extremes. A value near the top or bottom usually maps to noticeably different wording in the instruction, while values in the middle of the range often produce identical text. Test with the minimum and the maximum rather than nudging from 5 to 6 and hunting for a difference.
Does changing her appearance change her personality? No. Appearance parameters go to an image system that never receives your conversation, so they cannot affect wording. Any personality change you notice after a visual edit came from the app rewriting part of her description in the background at the same moment.
Why does she stop sounding custom after a long conversation? Because your own messages accumulate in the same window as the instruction, and by message 60 there is far more of your voice in there than of the description. Opening a fresh chat restores the balance immediately. Repeating one key behavioral line partway through a long conversation also helps.
A settings page is a sales document as much as a control panel, and the number of controls on it is a product decision rather than a measure of how configurable the character really is. The two fields that matter are unglamorous: a text box where you write sentences, and a switch that decides whether yesterday exists. Everything else is either downstream of those or decorative. Spend your effort accordingly, and check the results in a fresh chat, because a long one will tell you what you already put into it.
You do not have to take that table on faith. Pick a companion app such as Lona, spend your first ten minutes in the description field instead of the sliders, then look again a day later to see whether it remembered. That one test tells you more than the whole settings page.
Sources
- Vaswani, A. et al., “Attention Is All You Need,” 2017
- Sharma, M. et al., “Towards Understanding Sycophancy in Language Models,” 2023
