ChatGPT Projects vs Memory vs Custom Instructions: Which to Use and How to Set Them Up

You've told ChatGPT you're a finance analyst who wants short answers, three times this week. Yesterday it wrote you a motivational essay anyway. Meanwhile your side project's chats are scattered across the sidebar and every new one starts from zero.

Three features exist to stop this: Custom Instructions, Memory and Projects. They overlap enough that people turn on all three, forget which one is steering the answer, and then get confused. This post separates them, gives a setup I'd actually use, and covers the privacy settings that matter. Everything about behaviour is from OpenAI's help articles, read on 2026-10-08. OpenAI says memory features vary by plan, region and platform, so a setting I describe may be missing in your account.

What is the difference between Projects, Memory and Custom Instructions?

The rule I'd use: preferences go in Custom Instructions, jobs go in Projects, and Memory is a convenience you audit. Don't rely on Memory for anything that must be right, because by OpenAI's own description it decides what's relevant, and it "does not retain every detail".

What do Custom Instructions do, and what should go in them?

OpenAI describes them as direct guidance about what ChatGPT should know and how it should respond, applied immediately to all chats, including existing conversations. On web and desktop you find them under Settings, Personalization, with an Enable customization toggle. Updates only affect future behaviour; they don't rewrite old chats.

What belongs there: things that are true in every conversation and that you'd otherwise correct again and again. What doesn't: project facts, long documents, or anything that changes weekly.

Here's a version that fits the 1,500-character limit (roughly 600 characters, so there's room):

About me: I'm a finance analyst at a mid-sized logistics company. I use Excel
and SQL daily. I don't write production code.

How to answer:
- Give the answer first, then the reasoning in at most five bullets.
- If my question is ambiguous, ask one clarifying question before answering.
- Say "I'm not sure" when you aren't. Don't invent figures, sources or quotes.
- For spreadsheet help, state whether the formula is for Excel or Google Sheets.
- No motivational language or sign-offs.

Notes on why it's built this way. The instruction "ask one question if ambiguous" costs you a turn sometimes but beats a confident wrong answer. The "say you aren't sure" line doesn't make hallucinations disappear; it gives the model permission to hedge. And the list is short on purpose: the more rules, the more likely two collide. The system prompts lesson covers why short, non-conflicting instructions hold better.

Two things to know about data. Per OpenAI, your instructions aren't shared with people who view a shared link, and information from your use of custom instructions can be used to improve model performance unless you opt out. And if you use third-party plug-ins, OpenAI warns the model may pass relevant information from your instructions to the plug-in developer, so don't put anything in there you wouldn't tell them.

How does ChatGPT Memory work, and how do I keep it clean?

Memory is the feature people have the strongest feelings about, so read the controls rather than guess.

What OpenAI says it can draw on, depending on plan and region: past chats, saved memories, custom instructions, files in Library, and content from connected apps such as Gmail. There are two related things: saved memories, which are details you ask it to remember (or that it saves), stored separately from chat history, and reference chat history, where relevant details come from past conversations and can change over time. Settings, Personalization, Memory shows what your account offers; OpenAI lists controls such as Reference saved memories, Reference chat history and a memory summary.

A maintenance routine that takes ten minutes, once a month:

1. Open the memory summary (or ask ChatGPT "what do you remember about me?") and read it like a stranger would.
2. Correct what's stale. You can type a correction or highlight text and give one. "Don't mention this again" reduces future references but doesn't delete the source.
3. Delete what you don't want kept. This takes more than one step: deleting a chat doesn't necessarily delete a saved memory created from it. OpenAI's removal guidance is to delete the saved memory, the original chat, and anything else the information appears in (archived chats, Library files, connected apps).
4. Check the Sources line under answers. When ChatGPT uses personal context, Sources can show which past chat, memory, custom instruction or file contributed.

Know the limits: the summary is high-level and doesn't necessarily show everything Memory can reference. Turning Memory off doesn't delete your past chats, and if you turn it back on, it may build new remembered information from chats that are still in your history. Turning off Reference chat history schedules remembered information from past chats for deletion within 30 days.

If you want to read about how memory works inside agents generally, see AI agent memory architectures. It's a different layer than the ChatGPT consumer feature, but the trade-offs (what to store, when to forget) are the same.

When should I use a Project instead?

When the work has its own files and its own rules, and you come back to it. Think: weekly client reports, a thesis, a job search, a product spec, a trip. OpenAI's own examples are school (class materials in, study guide out), work (briefs and plans) and personal (vendors, budgets, timelines).

What a project holds: chats, uploaded files (PDFs, spreadsheets, docs, images or pasted text), project instructions, and optionally links to Google Drive files or folders and Slack channels as sources. You can also save a good ChatGPT response as a project source, which is how a project accumulates "decisions we've made" over time.

Project instructions are where the job-specific rules go. A template:

This project: weekly performance report for the Northwind account.
Audience: the account's marketing director, who reads on mobile.
Always: lead with the three numbers that changed most vs last week.
Format: 150 words max, then a table of the five main KPIs.
Source of truth: the latest file named "KPI export" in project files.
If numbers conflict between files, say so and ask which to trust.
Do not: speculate about causes unless I ask.

The "source of truth" line is the one most people skip. When a project has three versions of the same file, the model has no way to know which is current.

How instructions stack

Per the Projects help page, project instructions apply only inside that project and override your global custom instructions. In practice that means if your global instruction says "answer in five bullets" and your project says "write in full paragraphs", the project wins. On Enterprise and Edu, OpenAI's table shows custom instructions as not available in project chats at all, only project instructions. So for those accounts, repeat anything essential in the project instructions.

Practical limits and gotchas

• File caps: Free 5, Go and Plus 25, Edu, Pro, Business and Enterprise 40 per project, and only 10 files can be uploaded at once. If you hit the cap, OpenAI suggests removing old uploads, combining file data, or splitting into multiple projects.
• Google Drive in a project doesn't sync. OpenAI notes the Drive app doesn't sync when added within a project; it can search and access files but content isn't pre-synced for retrieval. If a file changes, re-upload it.
• Moving chats in: drag a chat onto the project or use Move to project. The chat then inherits the project's instructions and files. Chats created with a custom GPT can't be moved.
• Study mode doesn't apply to project conversations.
• Projects vs GPTs: OpenAI frames a GPT as a static, reusable custom assistant, and a project as an evolving context hub for recurring work. If you're building something for other people to use, read the custom GPT best practices instead.

If you use Claude too, Claude Projects work similarly, with differences in how files and instructions are handled.

What is project-only memory, and when should I turn it on?

Each project has a Memory setting, either Default memory or Project-only memory. Find it under the project's more options menu, Project settings, Memory, then Save. OpenAI notes changes can take a few hours.

With project-only memory:

• Your previously saved memories aren't referenced in project chats.
• Chats can reference other conversations in the same project.
• Chats can't reference conversations outside the project, and chats outside can't reference the project's.
• Switching an existing project to project-only removes information from that project from memory used outside it.

With default memory the behaviour depends on plan. On Enterprise and Edu, project chats stay contained anyway. On other plans, including Business, project chats can reference non-project conversations (unless the other project is project-only), and can be referenced from outside.

My rule: turn on project-only memory for anything sensitive or any work you don't want leaking into your personal chats (client work, health, finances, job search) and for anything where stale context would hurt. Leave default memory for projects where cross-pollination helps, like a general writing project where ChatGPT knowing your style is useful. There's no global switch; it's set per project. And you can't see a list of project memories. To stop it using a specific conversation, delete that chat or move it to another project.

What are the privacy implications?

Be honest with yourself about which of these you've checked.

Training. For personal accounts (Free, Plus, Pro), OpenAI says it may use information accessed from projects to train models if "Improve the model for everyone" is on. For Business, Enterprise, Edu and ChatGPT for Healthcare workspaces, OpenAI says content isn't used for training by default. If you don't want your chats used, that's a switch in Data controls, and it's worth checking now rather than after you've loaded a client's files.

Shared projects. Sharing is available on all plans, including Free, with a collaborator cap by plan (Free 5, Plus and Go 10, Pro 100, and workspace projects up to 100). Members can view the project's chats, files and member list, and any member can download the files. If an owner sets the link to "Anyone with a link", any logged-in ChatGPT user with the link can join. For personal accounts, shared project data may be used for model improvement only if the owner and every contributor have that setting on. Shared projects automatically use project-only memory and can't see a member's personal memories or custom instructions.

Temporary chat. For anything you'd rather not have remembered, use it. OpenAI says temporary chats don't create or update memories, even with personalization on. You choose Personalized or Unpersonalized before you start and can't change it later. Saving it converts it to a regular chat that then follows your normal settings. A temporary chat can't be added to a project. OpenAI's temporary chat help says a safety copy can be kept up to 30 days.

Deleting. Deleting a project permanently removes its chats, instructions and files stored only in the project; files saved separately in Library aren't deleted. OpenAI says deleted project data is removed from its systems within 30 days unless it must retain it for legal or security reasons.

Sensitive details. OpenAI itself says sensitive information can appear in Memory if you share it. Don't paste passwords, account numbers or medical details into any of these three features, and if you did, run the removal steps above.

A setup that works

1. Custom Instructions: the short block above. Five bullets, no project facts. (5 minutes)
2. Memory: switch it on if you want it, then do the first audit immediately. Put a monthly reminder in your calendar. (10 minutes)
3. One project per recurring job. Name it for the output ("Weekly Northwind report"), not the topic. Add instructions with audience, format and a source of truth. Upload only current files. (10 minutes each)
4. Project-only memory for work or personal-sensitive projects. Do it at creation, as the setting is easiest to reason about from the start.
5. Save good outputs back to the project using the message menu, so the project gets better with use.
6. Test it. Open a new chat in the project and ask: "What instructions and files are you working from?" If the answer doesn't match what you set up, fix the instructions before you rely on them. The context engineering overview explains why what's in the context window is what actually steers the answer.

Common problems and fixes

None of these features make ChatGPT smarter. They decide what it sees before you type. Spend your effort there, and keep the sensitive things out of all three.