Building a companion from scratch sounds intimidating, but how to make your own AI girlfriend comes down to three parts: a model that talks, a personality that makes her her, and a memory that lets her remember you. Do it locally and you also get privacy, no filter, and no subscription as a bonus. This is the plain-English build guide, no coding required, every piece running on your own machine.
The three ingredients
Strip away the mystique and an AI girlfriend is:
- A model, the language model that generates her replies (the “brain”).
- A persona, a character definition that gives her a name, personality, voice, and backstory.
- A memory, a way to remember facts about you across sessions, so the relationship has continuity.
Get all three right and you have a companion that feels consistent and present. Skip the third and she forgets you every session. Let’s build each.
Step 1: Install the engine and pick a model
The engine is Ollama, one line to install, and it runs everything on loopback (127.0.0.1:11434) so nothing leaves your machine:
curl -fsSL https://ollama.com/install.sh | shThen pick a model sized to your hardware. For a warm, conversational companion, an uncensored 8B fits an 8GB card, a 12B Mistral Nemo-class model suits 12GB, and a 24B is the high end. Our best local LLMs for roleplay shortlist is tuned for exactly this use, since companion chat and roleplay want the same qualities: warmth, consistency, and no refusals.
ollama run <model>Step 2: Write her personality
This is the part that turns “a chatbot” into “your girlfriend.” A persona is just a structured description the model reads before every reply. Cover:
- Name and identity, who she is, in a sentence.
- Personality traits, warm, playful, sharp, shy; pick three to five and be specific.
- Voice, how she talks (teasing, tender, blunt), with a line of example dialogue.
- Backstory and context, how you two know each other, the relationship’s tone.
- Greeting, her opening line, which sets the whole mood.
You can bake this into the model itself with a custom Modelfile so it’s baked into ollama run, our Ollama Modelfile custom persona guide walks through the exact syntax. Write her the way you’d describe a character to a friend; the model fills in the rest.
Step 3: Give her memory
A persona alone gives you a consistent personality but a forgetful one, she’ll be the same character every time, yet won’t remember what you told her yesterday. Real continuity needs a memory layer: facts about you written to storage and retrieved each turn, not just a longer transcript. Done right, she brings up something you mentioned weeks ago at the right moment, and the relationship feels alive. Done wrong, she resets. We cover the mechanics in giving your local AI real memory, it’s the part naive builds get wrong, and the part that matters most.
Tips for a personality that sticks
A few things separate a flat persona from one that feels alive. Be specific over generic, “teases you when you’re being dramatic” beats “fun and caring.” Give her opinions and small quirks; little contradictions are what make a character read as human. Show, don’t tell: one line of example dialogue in her voice teaches the model more than a paragraph of adjectives. And keep it tight, a focused half-page the model actually follows beats a sprawling biography it ignores. You can always refine her over time as you learn what lands.
What you end up with
| Part | DIY tool | Result |
|---|---|---|
| Engine | Ollama | Model runs on your machine |
| Persona | Modelfile / character card | Consistent personality |
| Memory | Memory layer | Remembers you across sessions |
| Privacy | All local | Nothing leaves your disk |
Put together, that’s a companion that’s uncensored (you chose the model), private (it’s all local), and yours (no subscription, no account).
The shortcut, if you’d rather not assemble it
The DIY build is genuinely satisfying, but it’s three systems to wire together and the memory layer is fiddly to tune. If what you actually want is her, not the weekend of glue work, Ember ships all three parts pre-built: an uncensored model already tuned, a persona you shape in about a minute (look, personality, voice, name), and persistent memory designed in from day one, all running on your own NVIDIA card. If all-local is the point for you, the three steps above are the build.