How to Make ChatGPT Remember Your Brand Voice
To make ChatGPT remember your brand voice, write the voice down as a short document with real example sentences. Then put it where ChatGPT reads it every time: a project’s instructions and files for brand work, or custom instructions if you only have one brand. Saved memory helps a little, but it’s too loose to hold a full voice.
After that, update the document whenever you correct the same mistake twice.
Why does ChatGPT forget your brand voice?
It usually doesn’t forget so much as never fully learn it. Each new chat starts from ChatGPT’s default style plus whatever context it has, such as your custom instructions and what you type. If your voice only lives in a chat from weeks ago, a new chat can’t count on it.
Three things make it worse:
- Long chats drift. The further you get from your original instructions, the more the default style creeps back.
- Vague descriptions. “Friendly but professional” describes almost every brand, so ChatGPT fills in the gaps with its own habits.
- Memory is a short list of facts. OpenAI’s Memory FAQ describes memory as details ChatGPT picks up across chats. That’s good for “I run a bakery in Lisbon.” It’s not built to hold a page of example sentences and a list of banned words.
What should a brand voice document include?
Keep it to a page. Long enough to be specific, short enough that it’s read in full every time. Here’s an example for a fictional brand, Juniper Lane, a small online shop selling houseplants.
| Section | What to write | Juniper Lane example |
|---|---|---|
| Who we are | One or two sentences | Juniper Lane sells hard-to-kill houseplants to people who’ve killed a few before. |
| Who we talk to | The reader in one line | Renters in small apartments, new to plants, a bit nervous about it. |
| Voice in three rules | Rules you can check | Short sentences. Gentle humor about plant deaths. Never make people feel dumb. |
| Words we use | Specific vocabulary | “Plant parent,” “bright indirect light,” “thirsty” |
| Words we avoid | Banned words and phrases | “Easy-care” (overused), “guaranteed,” exclamation marks in subject lines |
| Example sentences | Real lines from past posts and emails | “Your pothos is dramatic, not dying. Give it a drink and a day.” |
| Formatting | Platform habits | Instagram captions under 60 words. Emails open with one short line. |
The example sentences do the most work. Pull them from posts and emails that performed well or that you’re proud of. If you’re starting from nothing, write five yourself and add more as you edit drafts.
Where should you put your brand voice in ChatGPT?
You have three main places. Each fits a different situation.
| Where | Best for | Limits |
|---|---|---|
| Custom instructions | One brand, used in most of your chats | Applies everywhere, including chats that aren’t brand work; has a character limit |
| Project instructions and files | Brand work, or several brands | Only applies inside that project |
| Memory | Small stable facts about you | Not reliable for detailed rules; you can’t control exactly what’s stored |
Custom instructions
If ChatGPT is mostly for one brand, put a short version of the voice in custom instructions: who you write for, three voice rules and your banned words. Keep the full example list elsewhere, since space is limited.
Projects
For real brand work, a project is the better home. OpenAI’s help center says project instructions apply only inside that project and override your global custom instructions. Put the voice rules in the project instructions and upload the full voice document, with example sentences, as a project file. Do all writing for that brand inside the project. If you run more than one brand, make one project each; the guide to managing multiple brands with AI covers keeping them apart.
Memory
Use memory for small facts that don’t change, like your business name and your role. You can ask ChatGPT to remember something, and review what it saved under Settings, Personalization, Memory. Don’t rely on it for the voice itself. Some ChatGPT plans manage memory automatically, keeping the most relevant details prioritized, which means a rule you saved months ago may not be the one it reaches for.
How do you stop the voice drifting over time?
Getting the voice in is half the job. Keeping it is the other half.
Start every piece with a voice check. Ask ChatGPT to read the voice document and list the three rules it will follow before drafting. If the list is off, fix it before you get a full draft.
Correct with rules, not rewrites. When a draft misses, don’t just fix it yourself. Tell ChatGPT the rule it broke (“we don’t use exclamation marks in subject lines”) so the correction is explicit.
Update the document weekly. Any correction you made twice goes into the voice document. Any line you loved goes into the example sentences. This is the step that makes the voice get better over time instead of slowly drifting.
Start fresh chats often. A new chat in the project reads the instructions and files fresh. A very long chat is mostly reading its own earlier replies.
What if you use other AIs besides ChatGPT?
Then the voice needs to live somewhere all of them can read. Claude has its own projects and Cursor has its own settings. Neither sees ChatGPT’s. If you paste the voice into each one, you now have three copies to keep in sync, and they will drift apart. Claude memory vs ChatGPT memory explains why each app’s memory stays inside that app, and Claude Projects vs ChatGPT Projects compares the two project systems.
The cleaner setup is one source of truth that every AI reads, connected over MCP, the protocol ChatGPT and Claude both support for outside tools. What is MCP explains how that works.
FAQ
Can ChatGPT learn my writing style from examples?
Yes, and examples work better than descriptions. Give it a set of real sentences in your voice and ask it to match them. Keep those examples in a project file so every chat has them.
Why does ChatGPT keep using words I told it not to use?
Usually the instruction is buried in an old chat or a long document. Put banned words in a short list in your project instructions, and remind it at the start of a draft.
Should I use memory or custom instructions for brand voice?
Custom instructions or project instructions. Memory is for small facts; it isn’t designed to hold detailed voice rules, and you don’t control exactly what it keeps.
How long should a brand voice document be?
About one page. Long enough for real examples and banned words, short enough to read in full every time.
Where Valta fits
Valta is a hosted vault of Markdown notes that ChatGPT, Claude and other AIs connect to over MCP. Each vault has a profile (_vault.md) with its voice and rules, and connected AIs read it before writing for that vault. So the Juniper Lane voice lives in one place, and every connected AI reads the same current version. When you add a new rule on Friday, every AI has it on Monday.
Your AI does the writing; Valta keeps the voice, the example sentences and your brand notes in one vault, with a separate vault per brand on Pro. Create a free vault.