How to Set Up AI Custom Instructions and Memory (2026)

You are reading Chapter 11 of 36 in the AI Bootcamp, AI Central's free course for professionals who want to use AI in their own work. The chapters build on each other, so if this is your first one, Chapter 1 is the place to start. All of them sit on the AI Bootcamp page, and three more arrive every week.

AI custom instructions are a short standing brief you type into settings once, which your assistant reads before every new chat. Memory is the separate thing: notes the assistant writes about you on its own, from what you happened to mention. Setting up the first and checking the second is what stops you re-explaining your job at the top of every conversation.

This guide covers the four parts of a standing instruction worth writing, the container that holds the context for one recurring job, and the quarterly review that keeps both honest.

We publish AI systems like this for 300,000+ senior professionals at AI Central, and more of them live in the AI Central Library.

You open a new chat and type some version of the same four sentences you typed yesterday: this is my role, this is who the output is for, this is roughly how long it should be, please stop opening with a compliment. Then you get to the actual question.

Paying that twenty times a week is the quiet reason people stop using the tool. Not because it failed them, but because the run-up got tiring.

The description of your own voice you worked out in Chapter 10 is the raw material for most of what you are about to save. If you have not written that description yet, do it first and come back.

Two separate features exist to end the re-explaining, and almost everyone confuses them.

The first is a standing instruction: a short brief you type into settings once, which the assistant reads before every new chat. You wrote it, you can see it, and it does not change until you change it. All four assistants have this, on every plan including the free ones. ChatGPT calls it custom instructions, Claude calls it Instructions for Claude, Gemini calls it Instructions for Gemini, Copilot calls it custom instructions.

The second is memory. The assistant reads what you say across your chats, decides for itself which facts are worth keeping, and writes them down in its own words. You did not choose the wording and you were not asked. ChatGPT, Claude, Gemini and Copilot all do this now, and Claude opened it to free accounts in March 2026, so the free tier is not an exemption.

One caveat if you use Google: on Gemini both features need a personal Google account, and neither works when you are signed in with a work, school or supervised one.

The difference shows up at the moment things go wrong. A bad standing instruction is a line of text in a settings box: you open it, you read it, you delete the line. A bad memory is a sentence you never wrote, sitting in a screen you have probably never opened, shaping answers you assumed were neutral.

Before you write anything, go and look at what yours already holds. Open settings, find personalization, and read the memory it has built. Most people are surprised at least once.

Most people open the custom instructions box, type "be concise and professional", and get nothing back for it. Concise and professional mean different things to a lawyer and a designer.

The box is small and the instinct is to fill it with adjectives. Adjectives are the part an assistant is worst at interpreting, because "professional" is not an observable property of a sentence. Length is. Structure is. A banned opening line is.

Four parts, in this order.

Your role, your organisation or sector, and which English you write in.

The audience, their seniority, and what they already know, so it stops explaining your own industry back to you.

A default length, whether you want the answer before the reasoning, and whether you want prose or bullets.

The standing list of things you do not want, ever. This is the part most people leave out and the part that changes the output most. The flattering opener, the padded close, the bulleted list of everything: those are default habits you can switch off permanently, in one line. Nobody does it, because it does not occur to them that the defaults are negotiable.

A few hundred words is the room you get, so this is a brief and not a biography. If you want the underlying rules for writing an instruction an assistant can act on, the 26 principles of prompt engineering is the longer version.

Where to type it: ChatGPT under Settings then Personalization, Claude under Settings then the Instructions for Claude field on your profile, Gemini under Settings and help then Personal Intelligence then Instructions for Gemini, Copilot under Settings then Personalization. In all four the exact menu wording has moved at least once in the past year, so look for the word personalization and you will find it.

A standing instruction fires on every chat you ever open. So the moment you put something in it that is true of only one client, one product or one quarter, every unrelated conversation gets slightly worse.

The fix is a container. All four assistants now have one: a named workspace that holds a set of chats, a set of files, and its own instructions that apply inside it and nowhere else.

ChatGPT and Claude both call it a project. ChatGPT projects take files, project-only instructions, and an option to keep that project's memory separate from your account, and they work on free with a smaller file cap. Claude projects have their own instructions plus an uploaded knowledge base, and free accounts get up to five. Gemini calls its version a Gem, with fixed instructions plus reference files, free included, though your Instructions for Gemini do not apply inside a Gem. Copilot calls its version a notebook, gathering the files, pages and chats behind one topic, with availability that widened through 2026, so check what your licence includes.

Pick one piece of recurring work to start with, and be strict about recurring. The board pack you assemble every month. The weekly commercial update. The proposal format you have now written fourteen times.

Then put three things in it: the files that work always refers back to, the instructions specific to that job, and the chats about it. The files part is the Chapter 7 habit made permanent instead of re-uploaded every time.

There is a second reason to use one, and it is separation rather than convenience. ChatGPT lets you set a project to draw only on what is inside it, keeping your account-wide memory out and keeping anything learned inside from leaking back. Claude gives each project its own memory space by default. If you work across clients who should never colour each other, that isolation is the feature, not the file storage.

What you are doing here is personal setup: your assistant, configured for your work. Packaging a repeated job into something a colleague can open and use is a different exercise with different requirements, and it is Chapter 32.

Set up exactly one project today. Not six. The one you would otherwise re-explain on Monday.

Context that carries between chats without being asked for also carries mistakes between chats without being asked for, and it does it silently. There are three ways it goes bad.

A wrong fact, recorded once and repeated for months. Mention in passing that you are launching in March, and an assistant with memory may write that down as a fact about you. Six months later, long after the date moved, it is still shaping answers on that basis. There is no error message for a stale belief, and the output does not look wrong, it just looks slightly off in a way you attribute to the model.

Context that travels. Something you mentioned about your health, your family or your job hunt surfacing in a draft for work. One client's situation informing a reply about another. That is the feature working as built, in a context you did not have in mind when you fed it. The escape hatch is the single-chat mode every assistant now has, called temporary chat in ChatGPT and Gemini and incognito chat in Claude, which leaves no trace in history and feeds nothing forward.

A stale preference. A rule that was right in March and is wrong now, still applied to every answer.

Chapter 9 gave you a rule about what never goes in the box. Extend it to the settings box. A standing instruction and a memory both persist, and on a company account an administrator can search and export them, so anything too sensitive for a chat is too sensitive for your custom instructions.

One more thing worth knowing without losing sleep over it. Because assistants can read documents and web pages, security researchers have shown it is possible to hide instructions inside a file or a page that get written into an assistant's memory and then act on every later conversation. Specific cases have been demonstrated and patched. It is a reason to read your memory occasionally, not a reason to turn everything off.

The habit that covers all of this costs ten minutes a quarter. Open the memory screen and read it as if someone else wrote it, because someone else did. Delete anything out of date, correct anything wrong, and check your standing instruction still matches the job you actually have now. Clearing your chat history is a separate action: deleting the conversations does not delete what was learned from them.

Put the review in your diary now, dated three months out. It will not happen otherwise.

Before: every new chat opens with four sentences of throat-clearing about who you are and what you want, twenty times a week, and you have never opened the screen where your assistant keeps its notes about you.

After: those four sentences are saved once in settings, one recurring job has its own container with its files already in it, and you have read what the assistant decided to remember and deleted the parts that were wrong. The chapter's three exercises take about fifteen minutes.

Two things carry into Chapter 12: your standing instruction, saved in settings, and the one project you set up, with its files already in it. Chapter 12 closes AI 101 by reducing the first eleven chapters to the five uses you will genuinely repeat, and it assumes that setup exists.

If you want a one-page reference to keep beside you while you write the instruction, start with our free AI cheat sheets. For the plans and privacy settings behind each of these four accounts, read our setup guide for ChatGPT, Claude and Copilot, and the rest of the syllabus lives in the AI Central Library.

A short standing brief you type into settings once, which ChatGPT reads before every new chat. Claude calls the same thing Instructions for Claude, Gemini calls it Instructions for Gemini, and Copilot also calls it custom instructions. All four have it on every plan, including the free ones.

In most setups, yes. ChatGPT, Claude, Gemini and Copilot all keep running notes about you now, and Claude opened memory to free accounts in March 2026. Assume it is running unless you or your IT administrator switched it off. On Gemini, both memory and instructions need a personal Google account and do not work on a work, school or supervised one.

ChatGPT: Settings, then Personalization. Claude: Settings, then the Instructions for Claude field on your profile. Gemini: Settings and help, then Personal Intelligence, then Instructions for Gemini. Copilot: Settings, then Personalization. The exact menu wording has moved at least once in the past year in all four, so look for the word personalization.

A standing instruction applies to every chat you open. A project applies to one job. If a piece of context is true of you in every chat it goes in the instruction; if it is true only of one recurring piece of work it goes in a project, a Gem or a notebook; if it is true only of today, type it in the chat.

No. Deleting a conversation and deleting the memory it produced are two separate actions. If you do not want a conversation carried forward at all, use temporary chat in ChatGPT and Gemini, or incognito chat in Claude.