What Can My AI See?

Besides the chat you have open, ChatGPT can use things saved from earlier. Which ones depends on your settings and your plan.

Depending on your plan and region, Memory can draw on several things. These are past chats, saved memories, custom instructions, files saved in Library (where ChatGPT keeps files you save), and content from connectors (links to other services, such as Gmail). Source

Each term is explained below, with where to find it in Settings.

Chat or Work: what is the toggle?

Chat is for quick, everyday questions. Work is for longer jobs with several steps, such as making a document, spreadsheet or presentation. Source

• Chat: ask questions, search, brainstorm, or get quick conversational help. Source
• Work: research a topic, analyse information, or create a document, spreadsheet, presentation, report or Site. Source
• Chat and Work chats appear together in your Recents list. Chat is always available. Source
• In the browser and the phone app, Work runs in the cloud, which means on OpenAI's computers rather than yours. It cannot directly access files on your computer. Source
• Free and Go Work is rolling out to paid plans except Free and Go, so you may not see the toggle. Source
• Plus and Pro Work is available for eligible paid plans. Source
• Business, Enterprise and Edu Whether Work can use local files depends on your plan and your workspace's settings. Source
• Unverified Whether Work uses your memory and custom instructions the same way Chat does. The pages we read don't say, so check your Settings.

The things ChatGPT can see

The current chat

The conversation you have open right now.

Memory

Memory is how ChatGPT carries details from one chat to the next.

It does not keep every detail from every conversation. Source

The list you can read in Settings is a summary. It does not necessarily include every detail or source that Memory can use. Source

Find it at Settings > Personalization > Memory. Source

Chat history

Your chats stay saved until you delete them or a workplace rule for how long chats are kept removes them. Archiving a chat does not change that. Source

Memory can use your past chats. A setting called Reference chat history is part of this. If you turn it off, the help page says information remembered from past chats is scheduled for deletion within 30 days, and the original chats stay in your history. Source

Deleting a chat does not necessarily delete a separate saved memory made from that chat. Source

Projects

A project is a space for one topic, with its own chats, files and instructions.

Project instructions apply only inside that project and override the custom instructions you set for all chats. Source

Each project uses either default memory or project-only memory. With project-only memory, chats can refer to other conversations in the same project but not outside it. Source

Work is not available in projects that use project-only memory. Source

Uploaded files

Files you upload or create can be saved separately from the chat that used them, in Library. Deleting a chat does not delete a file that remains saved in Library. Source

Files added to a project are kept until the project or the file is deleted, or a workplace rule for how long files are kept removes them. Source

Pages disagree The release notes say a newer feature called Space replaces Library for some accounts. The older Library help page doesn't mention Space. Your screen may say either, so look at what it shows. Source Source

Connectors

A connector links ChatGPT to another service, such as Gmail. ChatGPT's help pages call them connected apps, and the Settings screen is called Plugins.

Find them at Settings > Plugins. Older help pages say Apps or Plugins, so look for either. Source

Connector permissions are separate from your general settings, such as Memory, personalisation and how long chats are kept. Source

To remove a connector's access, disconnect it. Disconnecting stops future access and does not delete past conversations that already used the connector's content. Source Source

Custom instructions

Direct guidance you write about what ChatGPT should know and how it should respond. Source

Custom instructions are available on all plans, on the web, desktop, iPhone and Android apps. Source

Find them at Settings > Personalization > Custom Instructions. On the phone app, look for Customize ChatGPT in Settings. There is an Enable customization switch. Source

Pages disagree The help page says changes apply immediately to all chats, including existing ones. Its FAQ says they show up only in future conversations. Start a new chat after you change them, and check the result. Source

Free and Go You can save up to 1,500 characters. Source

Plus, Pro, Business, Enterprise and Edu You can save up to 5,000 characters. Source

Temporary Chat

A chat that stays out of your history and is not used to improve OpenAI's models while it remains temporary. A later section explains how to use it. Source

The training setting

A setting called Improve the model for everyone controls whether OpenAI may use your conversations to train its models.

Free, Plus and Pro OpenAI says it may use your content to train its models. The page doesn't name Go. Source

Unverified The pages we read don't say whether the switch starts on or off, so check it on your screen. Source

When it is off, your new conversations are not used to train OpenAI models, and they can still appear in your chat history. If you give a thumbs up or thumbs down, the whole conversation may be used for training. Source

Find it at Settings > Data controls, then Improve the model for everyone. On the phone app, open the sidebar, select your profile icon, then Settings. Source

Business, Enterprise and Edu By default, OpenAI does not use content from these workspaces to train its models. Source

In the browser, Work cannot reach files on your computer. In the desktop app, Work can open a local folder you choose and use the files in it, with your permission. Source

A local folder is a folder stored on your own computer, such as one inside Documents. It is different from a file you upload to a chat.

Browser or desktop app

• Browser and phone app: Work cannot directly access files on your computer. You can still upload files to a chat. Source
• Desktop app: the new ChatGPT desktop app includes Chat and Work. Choose ChatGPT from the top-left menu, then choose Work with the toggle at the top of the page. Source Source
• Opening a folder: in Work, you can open a local folder. OpenAI says to open a folder or project and grant access only to the files the task needs. Source
• The desktop app can use local files and desktop apps with your permission, when your plan and workspace allow it. Source
• Free and Go Work is not rolling out to Free or Go, so this section may not apply to you yet. Source
• You may also see Codex. It works with local folders, repositories, terminals and developer tools. This guide does not cover it. Source

What it can read and change

• It can use files in the folder with your permission. Local chats run on your computer. Source
• Messages and task context may still be stored in the cloud, even when the work runs on your computer. Source
• Unverified We don't know whether Work can change or delete files in the folder, or what it can reach outside it. The pages we read don't say. Treat the folder as something ChatGPT might change, and keep a copy of anything you can't replace.

A safe first step Our suggestion

1. In File Explorer (Windows) or Finder (Mac), make one new, empty folder for one project. Name it for the project, for example "Thesis reading".
2. Copy into it only the files that project needs. Leave your originals where they are.
3. In the desktop app, open Work and open that folder.
4. Ask for something small first, such as a summary of one file. Look at what it did before you ask for more.

Avoid opening your whole Documents folder. OpenAI's own advice is to grant access only to the files the task needs. Source

For more on staying safe with AI agents (programs that carry out tasks on your computer), see Clank Safely.

Start a Temporary Chat and choose Unpersonalized before you send your first message. Unpersonalized does not use memory, custom instructions or connectors. Source

Open a new chat, select Temporary, then choose Personalized or Unpersonalized. You cannot change the choice after the conversation starts. Source

What a Temporary Chat does

• It stays out of your chat history. It is not used to improve OpenAI models while it stays temporary, whichever choice you make. Source
• It does not create or update memories. Source
• If you do not save it, you cannot reopen it from history after you leave. Saving turns it into a regular chat. Source
• OpenAI may keep a copy for up to 30 days for safety. Source
• Eligible files from a temporary chat may be saved to Library when you save the chat. Source
• Safety features may still use limited, safety-relevant context in rare, high-risk situations to help ChatGPT respond more safely. Source
• Enterprise For eligible Enterprise workplace accounts, temporary chats stay available to the workplace for 30 days through OpenAI's Compliance API (a tool for workplace record-keeping). Source

Or turn things off for a while

• Custom instructions: go to Settings > Personalization > Custom Instructions and turn Enable customization off. Turn it back on when you want it. Source
• Memory: go to Settings > Personalization > Memory. Turning Memory off does not delete your past chats. The page also describes a separate option, Delete and turn off memory, which deletes what is remembered. Choose that one only if you want it gone. Source
• Projects: inside a project, its own instructions override the custom instructions you set for all chats. For a clean answer, chat outside any project. Source
• Connectors: their permissions are separate from Memory and personalisation. To remove a connector's access, disconnect it. Source

Not on its own. Use it as a first look, then check Settings.

• OpenAI's Memory page does say: "You can also ask ChatGPT what it remembers about you." Source
• But even the memory summary in Settings "does not necessarily include every detail or source that Memory can reference". It may leave out details that are less relevant, unsuitable for the summary, or taken from sources that are not shown one by one. Source
• Unverified No page we read says that ChatGPT's answer in chat is complete or accurate. Source
• Better: read the summary at Settings > Personalization > Memory. It is the place OpenAI describes for reviewing what is remembered, though it may not show everything. Source
• The release notes (5 May 2026) say you can tap the Sources button below a response to see the saved memories, past chats and custom instructions it used. Source
• If something remembered is wrong, you can tell ChatGPT it is incorrect or outdated. Source

Your own screen is the final word. If it differs from this page, go with what you see.

Open ChatGPT in another tab or window and work down the list. Ticks stay on this page only and disappear when you leave it.

• Chat or Work. Look at the top of a new chat. Do you see Chat and Work? Which one are you in? On the phone app, use the dropdown at the top of the screen. On the desktop app, choose ChatGPT from the top-left menu first. Source
• Memory. Open Settings > Personalization > Memory. Is it on? Read the summary, and remember it may not show everything. Source
• Reference chat history. On the Memory screen, look for this setting. The help page names it but not exactly where it sits, so look around that screen. Decide whether you want past chats used. Source
• Custom instructions. Open Settings > Personalization > Custom Instructions. Is Enable customization on? Read what is written there. Does it still describe you? Source
• Projects. For each project you use, open it, select ••• then Project settings, and look under Memory for Default memory or Project-only memory. Read its instructions too. Source
• Connectors. Open Settings > Plugins (older pages say Apps). Which connectors are set up? Disconnect any you don't need. Source
• Training. Open Settings > Data controls and look at Improve the model for everyone. Is it on or off? Source
• Temporary Chat. Open a new chat and select Temporary. Does it show Personalized or Unpersonalized before you send anything? Source
• Saved files. Open Library from the left sidebar and look at what is saved. Your screen may say Space instead. Older page Source Source
• Folder (desktop app only). If you use Work with a local folder, check which folder is open. It should be the one you made for that project. Our suggestion Source
• Ask your admin. Which connectors are on, whether Work and local files are allowed on your plan, and how long chats are kept. Business, Enterprise and Edu Source Source Source

Besides the chat you have open, Claude can use memory, your past chats, projects, files you upload, connectors (links to other services, such as Gmail) and the instructions you set. Which ones depends on your settings and your plan. Details and sources are below.

One conversation, or Cowork (where Claude works through a task for you)?

In Claude's new layout there is one conversation and no mode to choose. If your message box still shows Chat and Cowork options, you don't have the new layout yet. Source

• Only the desktop app gives the full Cowork experience, where Claude can also use your local files and browser. Source
• Claude reaches your local files, the built-in browser and computer use (where Claude clicks and types on your screen) only while the desktop app is open. Source
• Free Cowork is for paid plans only. You may not see it. Source
• Pro and Max Cowork runs on desktop, web, mobile and in the Claude in Chrome side panel. From 6 Oct 2026, new Cowork tasks on Pro and Max run in the cloud. Source
• Team and Enterprise Cowork is available. Computer use, where Claude clicks and types on your screen, is not available at this time. Source Source

The things Claude can see

The current chat

The conversation you have open right now.

Memory

Memory is how Claude carries details from one chat to the next.

Free, Pro and Max Memory is on by default on the web, in Claude Desktop and in Claude Mobile. Source

Team and Enterprise Owners control whether memory is available, and it is off for each member until they turn it on. Source

Claude saves memory as a set of individual topics as you chat. Everything it remembers is listed under Topics. Source

Find it at Settings > Memory. Older accounts use Settings > Capabilities and View and edit your memory. Source Source

By default Claude does not store topics about personal or sensitive subjects. You can change that with Include sensitive topics in memory. Source

Cowork sessions that run locally on your computer do not use memory. Source

Chat history

Your chats are listed under Chats and tasks in the left sidebar. A switch called Search and reference chats, at Settings > Memory, controls whether Claude searches past chats. Source Source

A deleted chat leaves your history immediately and is deleted from Anthropic's storage within 30 days. Source

Team and Enterprise Anthropic keeps your chats in the product to give you a consistent experience. A deleted chat leaves your history immediately and is deleted from Anthropic's storage within 30 days. Source

Projects

A project is a space for one topic, with its own files, instructions and memory.

Files and instructions you add to a project are used across all chats in that project. Source

Project instructions apply only to chats inside that project. Source

Each project has its own memory. To keep it fully separate, turn off Use account memory under Memory on the project's page. Source Source

Pages disagree One help page says free accounts can have up to five projects. Another marks project instructions "paid plans only". Check what your screen offers. Source Source

Team and Enterprise Projects can be shared with others. Chats inside a shared project stay private to the person who wrote them. Source

Uploaded files

For documents, Claude reads the text, not images inside them. PDFs of 100 pages or fewer are the exception: it also reads visuals, such as images, charts and graphics. Source

Files in a project's Files section stay available across conversations. Source

Unverified How long uploaded files are kept on consumer plans, and who can see them. We could not find an official answer.

Connectors

A connector lets Claude reach another service, such as Gmail, read your data there and take actions.

Claude uses the same permissions you have in the connected service. Source

The Gmail connector can search and read email, and send, reply to and forward it. Source

By default, each action Claude takes on your behalf needs your explicit approval. Source

Pages disagree The research page says Research (a Claude mode that runs several searches to answer a question) can use your connectors automatically, without asking each time. Treat Research as a time when connectors may run without a prompt. Source

Find connectors at Customize > Connectors. You can switch them on for one conversation only. Source

Data retrieved through connectors is stored on Anthropic's servers, and you can delete it by deleting the chat. Source

Free You can add one custom connector. Source

Instructions for Claude

What you write under Instructions for Claude applies to all of your conversations. Source

Click your initials in the lower left corner, choose Settings, then find Instructions for Claude. Source

Claude app skills are separate. They add specific behaviours, such as how Claude formats its replies. Source

Incognito chats

Temporary conversations that are not saved to your chat history or to Claude's memory. A later section explains them. Source

The training setting

A switch called Help Improve our AI models decides whether your chats can be used to improve Claude. Some other pages use a different name for it. Source

Unverified The pricing page lists model training as "Opt-out" for Free, Pro and Max. Opt-out usually means your chats are used unless you switch it off. The announcement of the choice says people were given a choice but not which way the default goes. The privacy help page gives no default either. We can't tell you whether yours starts on or off, so check the switch. Source Source Source

When it is on, Anthropic uses your chats and coding sessions, including to improve its models. When it is off, it will not use your previous or new chats for future model training, apart from the cases in the next paragraph. Source Source

The switch is not the only route. The same privacy page says Anthropic may also use a conversation that is flagged for safety review, including to train models its Safeguards team uses. It also uses chats where you explicitly opt in to training in some other way, such as joining its Trusted Tester Program. Source

Turning it off does not undo training that has already started or finished. Source

If the switch is on, Anthropic may keep your chats in a form that doesn't identify you for up to five years for training. Feedback you send about a reply is kept for five years. Source

Find it at Settings > Privacy, under Help Improve our AI models. Source

Team and Enterprise Anthropic does not train its models on your content by default. Source

Incognito chats are not used to improve Claude, even when the switch is on. Source

Claude reaches files on your computer only through the desktop app, only in folders you have connected, and only while the app is open. Source

A local folder is a folder stored on your own computer, such as one inside Documents. It is different from a file you upload to a chat.

Browser or desktop app

• Browser and phone app: Cowork is available, but only the desktop app gives the full experience, including local files and the browser. Source
• Desktop app: Claude can read and change your local files without you uploading or downloading them. Where those files are processed is covered under "What it can read and change" below. You need the Claude desktop app for Windows or macOS, open and connected. The page says Cowork needs the latest version of Claude for Windows. Source
• Some connectors run on your own computer, and those work only in the desktop app. Source
• Free Cowork is for paid plans only, so this section may not apply to you. Source

What it can read and change

• Claude can only read and write files in folders you have connected. Source
• Claude asks your permission before permanently deleting any file, in every mode. Source
• The default permission mode is Manual, where Claude asks before it takes actions. You can change the mode at any time using the permission mode control in the chat box. Source Source
• In Skip all approvals, nothing checks what Claude does. The same page makes a separate promise about permanent deletion: Claude asks first in any mode. Leave Skip all approvals for later, or for never. Source
• Files Claude opens through the desktop app are still processed on Anthropic's servers, not only on your computer. Source
• Your connected folders are listed under Trusted folders. The working folder is listed under Storage folder. Both labels are in Settings. Source
• Folder instructions add project-specific context when you pick a local folder on desktop. Source
• Pro and Max Computer use, where Claude clicks and types on your screen, is in beta. It asks permission before each application. There is no wall (a sandbox) between Claude and your applications. Source
• Pro and Max Anthropic advises against giving computer use access to sensitive apps such as banking, healthcare or government. Source
• Unverified We could not find a full list of what the desktop app cannot read or change outside your connected folders, or the exact button names for choosing one. The pages we read don't say, so follow what your screen offers.

A safe first step Our suggestion

1. In File Explorer (Windows) or Finder (Mac), make one new, empty folder for one project. Name it for the project, for example "Thesis reading".
2. Copy into it only the files that project needs. Leave your originals where they are.
3. In the desktop app, connect that folder to a Cowork task.
4. Keep the permission mode on Manual. Ask for something low-risk first, such as a summary of one file, and read Claude's plan before you let it continue.

• Anthropic suggests a dedicated working folder rather than broad access. Source
• It advises avoiding files with sensitive information, such as financial documents, and keeping backups of important files. Source
• Its first-step advice is to start with low-risk tasks like summaries, and to review Claude's planned actions before it proceeds. Source Source

For more on staying safe with AI agents (programs that carry out tasks on your computer), see Clank Safely.

Start an incognito chat. It does not use Claude's memory and is not saved to your chat history. Your personal instructions and style settings still apply in it. Source

Open a new chat outside a project and click the ghost icon in the upper right corner. You will see a black border and an Incognito chat label in the upper left. Source

Unverified We found no official way to pause Instructions for Claude for a single chat.

What applies in an incognito chat

• Memory: not used, and incognito chats are not added to future memory. Claude does not pull from them when it searches past conversations. Source
• Your instructions: they still apply. The incognito page says your personal preferences and style settings apply, and calls them profile information. That page uses older names, so assume Instructions for Claude apply too. Source
• Projects: incognito is only available for chats outside projects, so project knowledge is not in play. Source
• Training: incognito chats are not used to improve Claude, even if the training switch is on. Source
• Kept for: the chat is not saved to your chat history, but the incognito page says Claude keeps it for 30 days by default. The general retention page does not mention incognito chats. Source
• Afterwards: once closed, an incognito chat cannot be reopened or turned into a regular chat. Source
• New layout: incognito chats open in the previous chat experience, so Claude can't create files or run code in them. Source Source
• Work plans Anthropic's page for business customers says incognito chats are deleted automatically within 30 days, unless they are flagged as breaking Anthropic's usage policy. Source

Or turn things off for a while

• Memory, one chat: click + in the message box, then click Memory to turn it off for this chat. Do it before you send the first message. Source
• Memory, all chats: use Pause in Settings > Memory. Claude keeps its existing memory but won't use it or make new memories. Source
• Careful: Reset permanently deletes all memories, including project memories. Pause is the gentler choice. Source
• Past chats: switch off Search and reference chats at Settings > Memory. Source
• Connectors: you can switch them on for a single conversation, so leave them off in a chat where you want a clean answer. Source
• Our suggestion For Instructions for Claude, copy the text somewhere safe, clear it in Settings, and paste it back later. Anthropic's pages don't describe this, so check the result.

Not on its own. Anthropic points to a list in Settings, so check that.

• Anthropic's memory page says: "Everything Claude remembers is listed under Topics." Source
• Its import and export page suggests asking Claude to "write out your memories of me verbatim". That is a suggestion for moving memory elsewhere. Source
• Unverified No Anthropic page we read says that Claude's reply to a question like this is complete or accurate. Source
• So read the list at Settings > Memory. Older accounts use Settings > Capabilities and View and edit your memory. Select any topic to read it, edit it or delete it. Source Source

Your own screen is the final word. If it differs from this page, go with what you see.

Open Claude in another tab or window and work down the list. Ticks stay on this page only and disappear when you leave it.

• Layout. Look at your message box. Does it show Chat and Cowork options, or one conversation with no mode to choose? Source
• Memory. Open Settings > Memory. Is it on? Read the topics Claude has saved. Edit or delete any you don't want. Source
• Past chats. On the same screen, look at Search and reference chats. Is it on? Source
• Sensitive topics. On the same screen, look at Include sensitive topics in memory. Is it on? Source
• Instructions for Claude. Click your initials in the lower left corner, choose Settings, and read what is under Instructions for Claude. Does it still describe you? Source
• Projects. For each project, open its page and look under Memory. Is Use account memory on or off? Source
• Connectors. Open Customize > Connectors. Which are connected? Click one and choose Disconnect if you don't need it. Source Source
• Training. Open Settings > Privacy and look at Help Improve our AI models. Is it on or off? Source
• Incognito. Open a new chat outside a project. Can you see the ghost icon in the upper right corner? Source
• Folders (desktop app only). Look for Trusted folders (your connected folders) and Storage folder (the working folder) in Settings. The help page doesn't give the full path. Remove access to any folder you don't recognise, and check the permission mode in the chat box says Manual. Source Source
• Ask your admin. Whether memory is on, which connectors are enabled, and how long chats are kept. Team, Enterprise and Education Source Source Source