You can get most of the benefit of a “personal” AI without giving a company a profile of you. Write your preferences into a system prompt (called custom instructions in some apps), keep a short reusable note about yourself, and paste in context only when a task needs it. Cloud memory features do this automatically but store what they learn on the provider’s servers. A local model with a saved system prompt keeps the same information on your phone.
How do AI chatbots personalize their answers? #
There are three main mechanisms, and they differ in where your information lives.
| Mechanism | What it is | Where it’s stored |
|---|---|---|
| System prompt / custom instructions | Standing instructions the model reads before every message | Cloud apps: on their servers. Local apps: on your phone |
| Memory | Facts the chatbot saves about you from past chats | On the provider’s servers |
| Behavioral personalization | Apps learning from what you tap, read or skip | On-device for some features, cloud for others |
The first one does most of the work, and it’s the one you control completely.
What does cloud chatbot memory store? #
Memory features in cloud chatbots save details you mention, such as your job, your kids’ names, your diet or your writing style, and use them in later conversations. Some can also reference your past chats. It’s convenient, and it’s a running profile of you on someone else’s server.
If you use one, open its memory settings from time to time to see what it has saved, delete what you don’t want kept, or switch memory off. Temporary or incognito chat modes, where offered, skip memory for that conversation. Setting names change often, so check your provider’s help pages, such as OpenAI’s memory FAQ.
Build your own “about me” system prompt #
A system prompt you write yourself gives you the same personalization with nothing learned behind your back. Keep it short and include only what you’re comfortable storing.
About me: I'm a nurse working night shifts. I cook for two,
mostly vegetarian. I prefer metric units.
How to answer: Be direct. Use short paragraphs and bullet points.
Skip disclaimers unless something is genuinely risky.
If you're not sure, say so instead of guessing.That’s enough to make answers noticeably more useful. Leave out anything identifying that the model doesn’t need, such as your full name, employer or address.
Use a different system prompt for each job #
One big prompt that covers everything gets muddy. Separate chats with focused instructions work better, especially on smaller models:
- Recipe helper: “Suggest vegetarian dinners for two, 30 minutes or less, using what I list. Give quantities in grams.”
- Email editor: “Rewrite my drafts to be polite and brief. Keep my meaning. Never add information.”
- Study coach: “Quiz me on what I paste, one question at a time. Tell me if I’m right and why.”
- Code reviewer: “Review code I paste for bugs first, style second. Be terse.”
The recipe assistant doesn’t need to sound like the code reviewer. Few-shot examples inside these prompts help small models even more, as we show in zero-shot vs few-shot prompting.
Doing this on your phone, privately #
Personal LLM runs open models such as Qwen 3.5 and Gemma 4 entirely on your phone, and it’s built around this approach. Every chat can have its own system prompt, and a default system prompt under Settings, Behavior applies to every new chat, which is a good home for your “about me” note. One-tap presets (Creative, Balanced, Precise, Thinking) set the model’s style, and temperature, context size and other settings are yours to adjust.
All of it stays on the phone: the chats, the prompts and the settings. There’s no account and the company doesn’t run a server. If you want a copy, you can back up every chat and setting to a single JSON file that you keep, and restore it later.
What about apps that personalize on the device? #
Your phone already personalizes a lot without uploading much. Keyboards learn your words, photo apps build “memories” from your library, and suggestion features learn which apps you open in the morning. Much of that runs on the device. Where companies do learn from many users, some use federated learning, which sends model updates instead of raw data. Our federated learning explainer covers how much that protects.
What local personalization can’t do #
Be clear about the trade-offs:
- No automatic memory. A local chat app with a fixed model doesn’t learn about you between chats. You maintain the “about me” note yourself. Some people see that as a feature.
- Limited context. Everything in the system prompt uses part of the model’s working memory. Keep it to a short paragraph, not a biography.
- Smaller models. Personal instructions make a phone model more useful, not more knowledgeable.
For what’s worth keeping out of any chatbot, personalized or not, see is it safe to tell an AI chatbot personal things.
Frequently asked questions #
What are custom instructions? #
They’re standing instructions a chatbot reads before every message, such as how you like answers formatted or facts about your situation. Technically they’re a system prompt. Local AI apps let you set the same thing per chat.
Does ChatGPT memory store my data? #
Yes. When memory is on, ChatGPT saves details from your conversations on OpenAI’s servers and uses them later. You can view and delete saved memories, turn the feature off, or use a temporary chat that doesn’t use memory.
Can a local AI remember me between chats? #
Not automatically in most local apps, since the model doesn’t learn from your conversations. Put the details you want it to know in a default system prompt, and every new chat starts with them.
What should I put in a system prompt? #
Your preferences for tone and format, relevant context about your situation, and what the model should do when unsure. Keep it under about 100 words and leave out anything identifying you don’t need, like your full name or address.