Within one chat, yes, up to the limit of what it can hold at once. Across chats, only with memory on: ChatGPT keeps saved details, draws on past conversations and, since a June 2026 update, maintains a memory summary you can review. Temporary Chat and project-only memory keep work separate. On ChatGPT Business, the workspace owner controls memory.
For a business, memory is a trade. It saves you re-explaining who you are and how you write, but it can also carry one client's details into another client's work, keep facts that have gone out of date, and sit in a staff member's account after they've left. Knowing which kind of remembering you're seeing, and where each switch lives, lets you keep the upside and fence off the rest.
Three kinds of "remembering", and which one you're seeing
People use "ChatGPT remembers" to mean three different things. They behave differently and you control them in different places.
| What it is | What it covers | How long it lasts | Your control |
|---|---|---|---|
| The current chat | Everything said in this conversation, up to the model's working limit (its context window) | This conversation only | Start a new chat to clear it |
| Chat history | Saved transcripts of past conversations in the sidebar | Until you delete them; deleted chats are removed within 30 days | Delete or archive chats; use Temporary Chat to avoid saving |
| Memory | Details ChatGPT carries into new chats: things it has saved about you, plus context drawn from past conversations | Until you delete it or turn memory off | The memory controls in Settings |
The first one catches people out in long chats. If you've been working in one conversation all afternoon, details from the beginning can drop out of view or get less attention. When answers start ignoring something you said earlier, ask for a summary of the key facts, start a fresh chat and paste the summary in.
A typical case, as an illustration: an owner spends an afternoon in one chat drafting a tender response. Early on she corrects the team size from 12 to 8. Forty messages later, the staffing section confidently says "our team of 12". The correction hasn't been forgotten in any human sense; it has been outweighed by everything said since. The fix takes two minutes:
List the fixed facts of this tender in no more than 15 lines:
names, numbers, dates, prices and decisions we've agreed.
Mark anything I corrected during this chat as CORRECTED.She reads the list, which includes "Team size: 8 (CORRECTED from 12)", opens a new chat, pastes the list at the top and carries on. If the list itself gets a fact wrong, she has caught the problem before it reached the tender rather than after.
The third one is what changed most in 2026. In June, OpenAI moved memory to a background process that reviews your conversations and keeps its picture of you up to date, shown on a memory summary page, with the rollout reaching different plans at different times. The memory help page is candid about one limit: the summary won't include everything ChatGPT remembers, and the most reliable way to check whether it knows something is to ask it directly.
Where the switches are (as of September 2026)
Settings labels and availability shift by plan and change often, so treat this as a map and check what your own account shows.
- Memory on or off: in Settings, in the personalisation section. Depending on your plan you may see separate controls for saved memories and for using past chats, a link to your memory summary, and a way to manage or delete what's stored.
- Temporary Chat: start one from the chat menu. Nothing from it goes into memory or your history, and it isn't used for training; OpenAI can hold on to it for as long as 30 days for safety review before deleting it.
- Project-only memory: in a project's settings, under Memory. With it on, chats in the project can draw on each other but not on your general memories or chats outside the project, and the project's contents stay out of your general memory. It can take a few hours to take effect after you change it.
- ChatGPT Business: the workspace owner decides whether memory is available to members, under the workspace settings. Individuals can still turn their own memory off.
- Custom instructions: in the same personalisation section of Settings. Unlike memory, these are fixed notes you write yourself ("We're a six-person IT support firm; write in plain, friendly language; keep replies under 150 words") and they apply to every chat.
Custom instructions are often the better tool for stable business facts, because you wrote them and you can see exactly what they say. How to give AI your business context covers what to put there.
Filled in for an illustrative three-person landscaping business, custom instructions might read:
- "We design and build small gardens for homeowners; jobs run from $2,000 to $25,000."
- "Quotes always separate materials, labour and plants, and never include a total I haven't given you."
- "Write in plain, warm language, with no exclamation marks, and sign off as 'The team'."
- "Never mention a start or delivery date unless I give you one."
Every line is something the owner would otherwise type into dozens of chats, and none of it concerns a particular customer. That's the test for custom instructions: stable, business-wide, and safe for every conversation to see.
What memory is genuinely good for
Before the warnings, the upside. Memory earns its keep for the things you'd otherwise repeat in every chat: your role, the size and type of your business, the tone you like, the format you want reports in, the tools you use. An owner who has told ChatGPT once that they run a six-person firm and hate long emails gets shorter, better-aimed drafts without asking. If your work with ChatGPT is mostly your own admin and writing, leaving memory on and checking the summary now and then is a sensible default.
What memory gets wrong in business use
Memory is built for one person's life. Business use strains it in five predictable ways:
- Client crossover. Details from one client's chat turn up in a draft for another: the wrong product, the wrong contact name, the wrong setup.
- Stale facts. Last year's prices, a supplier you've dropped, a colleague who has left. Memory updates over time, but not on the day your business changes.
- Inferred preferences. One chat where you asked for a very formal letter can shape the tone of the next twenty.
- Shared logins. Several people using one personal account build one confused memory made of everyone's work. Give each person their own account.
- Work memories in personal accounts. If a staff member uses their own ChatGPT for work, the business details it remembers belong to their account and leave with them. That's one of several reasons customer data shouldn't go into personal plans.
The quickest way to find out which of these you have is to ask: "What do you remember about me and my work? List everything, one item per line." An illustrative reply from a bookkeeper's account:
- You run a small bookkeeping practice and prefer short, plain replies.
- Your hourly rate is $65.
- You're helping a café owner reconcile their card takings for March.
- You prefer formal letters that end "Yours sincerely".
- A colleague handles payroll queries on Thursdays.The first line is exactly what memory is for, so it stays. The rate is stale (in this illustration it rose to $75 in July), so correct it, or delete it and keep current rates in custom instructions. The café line is client work that belongs in a project or nowhere. The formal-letter preference came from one letter to a bank and has been making every client email stiff. And the colleague left in the spring. One kept, one corrected, three deleted, in about four minutes.
Worked example: an IT support firm sets memory up per client
Take an illustrative six-person IT support firm with 30 small business clients, all on ChatGPT Business. The trigger for change: a technician's draft email to a dental practice mentioned a firewall model that belongs to a different client entirely. Memory had quietly joined two conversations.
The owner changes the set-up in an afternoon:
- General memory stays on, for writing preferences and firm-wide facts only. Each technician checks their memory summary and deletes anything client-specific.
- Each of the ten largest clients gets a project with project-only memory, holding environment notes, standard fixes and past summaries, but never passwords or credentials. Setting up ten projects takes about two hours. How to use ChatGPT projects to keep client work separate covers the structure.
- Smaller clients go through Temporary Chat when a technician needs to paste an error log for a one-off question.
- A monthly ten-minute check: each technician asks ChatGPT, "What do you remember about our clients?", reads the answer and deletes anything that shouldn't be there.
The ongoing cost is about ten minutes per person a month. The crossover problem goes away, because client details only exist inside the project they belong to.
One snag turned up in the first week. A technician worked through a printer-driver error in Temporary Chat, closed it, and the next morning needed the same fix for a second client. Temporary chats don't appear in history, so the answer was gone. The habit that solved it: before closing a temporary chat, copy the useful part of the answer (never the pasted log) into the ticket notes, where the fix belonged anyway.
Choosing memory settings for your situation
| Situation | Suggested setting |
|---|---|
| Owner on Plus, one business, few confidential clients | Memory on; review the summary monthly; Temporary Chat for sensitive one-offs |
| Team on ChatGPT Business, several clients each | Memory on for preferences only; one project per major client with project-only memory |
| Work that's mostly confidential client material | Use of past chats off; rely on projects and custom instructions you control |
| Staff using personal accounts for work | Move them to a business plan; until then, memory off for work chats |
| Several people sharing one login | Stop sharing; one account each, then decide memory per person |
If you're deciding between a personal and a business plan partly because of these controls, ChatGPT Plus vs ChatGPT Business sets out the other differences. And if your team shares context more than it stores personal preferences, a shared project is usually cleaner than memory. Don't build a custom GPT for it: OpenAI is retiring them, and existing ones stop running on 11 December 2026. If your team uses more than one assistant, how ChatGPT projects compare with Claude Projects and Gemini Gems sets out the options.
Tools built on ChatGPT don't share its memory
Memory is a feature of the ChatGPT app, not of the underlying models. If you use a chatbot, helpdesk tool or automation that runs on OpenAI's models through the API, it doesn't know anything from your ChatGPT memory, and it doesn't remember one conversation in the next unless its maker built that in.
That cuts both ways. A website assistant built on the API won't suddenly mention something you told ChatGPT last week. But if the vendor has added its own "memory" or conversation history, that's stored on their systems under their terms, not OpenAI's. When you evaluate one of these tools, ask the vendor three questions: does it keep conversation history between sessions, where is that history stored and for how long, and can you delete a particular customer's history on request?
Answers vary more than you'd expect. An illustrative reply from a helpdesk tool's sales team: "Conversation history is kept for the life of your account so the assistant can refer to past tickets. It's stored on our own cloud servers. An admin can delete an individual customer's history at any time, and it's removed from backups within 35 days." That's workable: you know it remembers, where, and how to make it forget. Compare "We use enterprise-grade security and never sell your data", which answers none of the three questions. Ask again, in writing, until you get specifics.
The same applies to Zapier or Make automations with an AI step. Each run starts fresh. If an automation needs context, such as a customer's previous tickets, it has to be passed in on every run, which is usually what you want: you can see exactly what the model was given.
Clearing memory when someone leaves or a client asks
The main thing to know: deleting a chat doesn't necessarily delete a memory that was saved from it. The two are stored separately. To remove something properly:
- Ask ChatGPT what it remembers about the person or client, so you know what to look for.
- Delete the relevant saved memories or summary items in the memory settings.
- Delete the chats where the information came from.
- Allow time. Deletions take a while to work through, and OpenAI says deleted chats leave its systems within 30 days, barring security or legal reasons to keep them.
- Ask again a few days later to confirm it's gone.
Run through for the IT firm when a former client asks for its details to be removed from the firm's AI tools: the technician asks ChatGPT what it remembers about the client and gets two items, a firewall model and the office manager's name. Both are deleted, along with the client's project and the six chats that turn up when history is searched for the client's name. Four days later the same question gets "I don't have any saved details about that client." The reply to the client lists what was deleted and when, and says that OpenAI finishes removing deleted chats from its systems within 30 days. That's more honest than "all gone", and it's the answer the client can check against.
On ChatGPT Business, note one workspace-level effect: if the owner turns memory off for the whole workspace, members' existing saved memories are deleted. That's a clean reset if you need one, but warn the team first, because their saved preferences go too.
When a staff member leaves, deal with their account on the same day you remove their email access. On a business plan the workspace owner can remove them; on a personal plan, the account and everything it remembers stays with them, which is the strongest argument for not letting work happen there in the first place.
Questions about ChatGPT memory at work
Does memory mean my chats are used to train ChatGPT?
No, they're separate settings. Memory decides whether ChatGPT uses what it knows about you in later chats. Model training is controlled by the "Improve the model for everyone" switch in Data Controls on personal plans. Business and Enterprise workspaces don't train on your content by default, whether memory is on or off.
Can colleagues see my ChatGPT memories on a Business plan?
OpenAI's business memory guidance describes memories as tied to each individual account and not transferable to other users in the workspace. If your team needs shared context, such as client background or house style, put it in a shared project instead, where you can see exactly what's included and who has access. Custom GPTs used to be the other option, but OpenAI is retiring them: they stop running on 11 December 2026.
Do Claude and Gemini remember conversations too?
Yes. Both now have memory features that carry details between chats, and for many users the setting is on by default. Claude's memory is available on its free and paid plans, and Gemini can personalise answers using your past chats. Each has its own switch in settings, so check them the same way you'd check ChatGPT.
Further reads
- ChatGPT Temporary Chat: When Your Staff Should Use It — When staff should use Temporary Chat so work stays out of memory.
- How to Stop AI Tools Training on Your Business Data — The training switches, which are separate from memory.
- How to Set Up Claude Projects as a Shared Team Assistant — The Claude equivalent of a project-per-client set-up.
- How to Build a Shared Prompt Library for Your Team — Share good prompts across the team instead of relying on memory.
- Why AI Gives Different Answers Each Time, and How to Fix It — Why answers vary between chats, memory or not.
- How to Build a Company Knowledge Base AI Can Answer From — A deliberate home for business facts AI should always use.
- How to Use ChatGPT as a Business Adviser Without Being Misled — Prompts and checks that turn ChatGPT into a useful sparring partner for business decisions, and stop it flattering a bad idea into a plan.
- ChatGPT Prompts for Pet Sitters and Dog Walkers That Win Clients — Twelve ready-to-paste prompts for pet sitters and dog walkers, from profile bio to daily updates, each with what to feed in and what to check.
- ChatGPT Connectors: What They Can See in Your Drive and Inbox — What ChatGPT's Drive, Gmail, Outlook and SharePoint apps can read, where training and memory come in, and the admin switches that narrow access.
- When Does a Shared ChatGPT Plan Stop Being Enough for Your Team? — Eight signs one ChatGPT login has run its course, the sums for moving a six-person team to seats, and where ChatGPT Business itself tops out.
- Is ChatGPT Safe for Business Use? Risks, Settings and Plan Choice — The five real risks of using ChatGPT at work, the settings that fix most of them, and which plan fits a sole trader, a small team or a clinic.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: OpenAI help centre articles on memory, the business version of the memory FAQ, projects, Temporary Chat and data controls; OpenAI's June 2026 announcement on memory; Claude and Gemini help pages on memory (checked September 2026).