Memory vs instructions vs projects
How ChatGPT memory, custom instructions and projects differ, what each is for, whether projects share memory, and how to set ChatGPT up to answer the way you want.
Guide
What should your AI tools know about you? 5 things that change answers
Your AI tools need five things about you: who you are and what you do, how you write, who you work with, what you decided and why, and the work you repeat.
You ask ChatGPT for an email to a client and get something grammatically fine that sounds nothing like you. Longer than you'd write, stiffer than you'd write, and suggesting an idea the client turned down last week. Usually the tool isn't the problem. It just doesn't know enough about you.
By Musab Alrakhis · Last reviewed:
Here are five things that, once your tools know them, visibly change their answers. Each comes with an example you can adapt, followed by what you should keep out of any AI memory.
Why does knowing you change the answer?
An AI tool writes from what's in front of it: your request, the chat so far, and whatever it finds in memory. If all it has is a short request, it writes for "anyone". If it finds details about you and your work, it writes for you.
It's the difference between a new hire on day one and a colleague of a year. Both are capable; the second knows what you mean from half a sentence.
1. Who you are and what you do
Start with the basics: your role, your field, who you work for or with, and who you serve.
Example, in your own words:
"I'm an independent marketing consultant. I work with small and mid-sized food businesses in the Eastern Province. Most of my work is monthly content plans and performance reports."
Those three lines change a lot. The tool now knows examples should come from the food sector and that the reader owns a small business rather than running marketing at a large company.
2. How you write
Your writing style is what makes an answer sound like you. Think about length, tone, phrases you use and phrases you can't stand.
Example:
"I write short emails, three paragraphs at most. I get to the point without preamble. I don't use 'I hope this email finds you well' in work emails. With clients I write plain, friendly English."
To make it sharper, give your tool two or three emails you actually wrote and ask it to summarise your style and save the summary. Then read the summary and correct it.
If you write to some clients in Arabic and others in English, say so, and describe what fits each. Many people are more formal in one language than the other, and the tool won't know unless you tell it.
3. Who you work with, and what they prefer
Most work involves other people: clients, managers, colleagues, suppliers. Each has preferences that keep coming up.
Example:
"Northwind (client): wants the report before the 5th of each month, opening with a one-page summary. Dislikes complex charts. Main contact: Sara, operations lead."
Save that to the client's project and any tool will write next month's report the right way without a reminder.
Mind personal data here. Keep only what the work needs, such as a contact's name and role, and leave out personal phone numbers and other details you don't need.
4. What you decided, and why
Decisions and their reasons are the most valuable thing your tools can know. A tool that doesn't know you rejected an idea will suggest it again.
Example:
"We decided not to post on Snapchat for this client, because their audience is business owners who engage more on LinkedIn and X. Decided in the 3 October meeting."
Note the decision, the reason and the date. The reason helps the tool judge similar cases you didn't spell out. The date helps you tell whether the decision still stands.
5. The work you repeat, and how you like it done
Everyone has recurring work: a weekly report, meeting notes, a reply to a common question. Explain once how you want it done and you save yourself the explanation every time after.
Example:
"When I ask for meeting notes: start with decisions, then tasks with owner and due date, then open questions. One page maximum."
Another:
"When I ask for a reply to a pricing question: don't quote numbers, offer a short call, and say we send a written proposal after the call."
In Memory Bridge you can save these so all your tools do the work the same way.
What the answer looks like before and after
Take one short request: "Write to Northwind about this month's report running late."
Before the tool knows you: five paragraphs, a long greeting, generic apologies, a suggestion to "schedule a meeting to discuss", and an empty sign-off. Correct, but you'll rewrite most of it.
After: two paragraphs addressed to Sara by name, straight to the point, confirming the report will arrive before the 5th as usual and offering the one-page summary first. A quick read and it's ready to send.
The difference didn't come from a smarter model. It came from four saved lines about the client and your style.
This is an illustration, not a customer story.
Where should this information live?
Three places, each with limits.
Each tool's custom instructions
Good for a few stable facts, like items 1 and 2. But they belong to one tool: you write them in ChatGPT, write them again in Claude, and update each separately.
Each tool's built-in memory
It keeps some of what the tool judges important and stays inside that tool. The tool chooses what to keep, so the details you care about may not make it.
One memory every tool reads
In Memory Bridge this information is saved as plain-text notes inside projects, each with its source and date. Every tool you connect reads them, such as Claude, ChatGPT, Gemini and Cursor. To avoid writing it all from scratch, start with the introduction session: your tool asks about your work, your clients and how you like to work, and saves your answers as notes you review and correct.
What not to tell any AI tool
Whatever the tool or service, keep these out of memory:
- Passwords, access keys and verification codes.
- Card and bank account numbers.
- Other people's personal data your work doesn't need, such as ID numbers, home addresses or health information.
- Information your employer classifies as confidential, unless its policy allows it.
How do you know it's working?
Once the five are saved, run a simple test. Open a new chat and make a very short request, like "Write to Northwind about this month's report." If the reply is short, in your style, with the right date and the contact's name, the tool knows you. If it's generic, check what was saved and fill the gaps.
Review your notes about once a month. Work shifts, clients change, old decisions get reversed. A stale note can lead a tool to the wrong answer.
Your next step
Write the five things down for yourself today, even two lines each. Then save them somewhere every tool can read, so you never have to type them once per tool.
How this article was written
The Memory Bridge team drafted this article with help from AI tools and checked it against the live product and the sources linked above in October 2026. Menu names in other tools may change after that.
Common questions
What's the most important thing to tell an AI tool about myself?
Your role, your field and who you serve, then your writing style. Those two change answers more than anything else.
Should this go in custom instructions or memory?
A few stable facts suit custom instructions. Things that change with your work, like decisions and client preferences, suit a memory the tool updates and you review.
Can my tool learn my style from examples?
It can summarise your style from examples and save the summary. Review and correct it, because the tool may describe it wrongly.
What shouldn't I save in AI memory?
Passwords, access keys, card numbers, and other people's personal data your work doesn't need.
How often should I review what my tool knows about me?
About once a month, and whenever something important in your work changes.
Keep reading
Set up in 10 minutes
A practical walkthrough of your first minutes with Memory Bridge: connect one tool, introduce your work, test recall in another tool, review your notes.
An AI memory you can read and edit
Your AI memory is a set of clear notes inside projects. Read what was saved, correct it, go back to an earlier version, and download a copy to keep.
One memory. Every tool you use.
Pay for a month or a year. Nothing renews automatically.