Use Custom Instructions for stable preferences that should apply everywhere, Memory for evolving facts you want ChatGPT to carry forward on its own, and Projects for scoped work that should stay walled off from the rest. The three overlap, which is why people mix them up, but each was built to solve a different problem. The one-line rule: instructions are what you write, memory is what it writes, projects are where both stay sealed off.
Custom instructions: your standing rules
Custom instructions are a fixed field you fill in by hand. They tell ChatGPT how you want it to respond and what background to assume, and they apply to new chats across the board. Nothing here changes unless you change it. This is the right home for stable preferences: the language you want, the format you like, your level of expertise, the role you play. Because you control the text directly, custom instructions are predictable in a way the other two are not.
Memory: facts it carries on its own
Memory is the part that updates itself. ChatGPT notes facts as you chat, and you can also tell it to remember something directly. Then it reuses those facts in later conversations without you repeating them. The upside: it adapts as your situation changes without you lifting a finger. The catch is control. Memory decides what to keep, and the saved-memory store is small enough to fill up, so an offhand remember that can crowd out something you cared about. Treat it as helpful but not authoritative.
Projects: a walled workspace
Projects group related chats, files, and instructions into one workspace. A project can hold its own files and its own custom instructions that apply only inside it, and the chats you start there stay together. The point is isolation: work in a project does not bleed into your general chats, and your general context does not clutter the project. This makes projects the right tool for a specific client, a sensitive topic, or anything you want kept separate and self-contained.
Which to use: a quick decision
- Should this apply to almost every chat and rarely change? Use custom instructions.
- Is this an evolving fact you would rather not repeat? Let memory hold it.
- Does this work need its own files and to stay separate? Make it a project.
- Is it sensitive or one-off? Use a project or a temporary chat, not memory.
| Feature | Scope | Who controls it | Best for |
|---|---|---|---|
| Custom instructions | All new chats | You, by hand | Stable preferences |
| Memory | Across chats | ChatGPT, plus you | Evolving facts |
| Projects | Inside the project only | You, per project | Scoped or sensitive work |
Where they trip each other up
Most guides tell you to lean on memory. The opposite is usually right: memory is the least controllable of the three, so the more you care about a behavior, the more it belongs in custom instructions, not memory. Memory can quietly capture something that contradicts your custom instructions, and the model then has to reconcile the two. A casual instruction inside one chat can get saved as a memory and follow you into unrelated chats. Projects help by sealing context off, but anything memory saved before you started the project can still surface. The habit that keeps this clean: put durable rules in custom instructions, review memory now and then, and push anything sensitive into a project.
If ChatGPT keeps applying a preference you no longer want, check memory first, not custom instructions. Auto-saved memories are the usual culprit behind behavior you cannot find a setting for.
Beyond the three: durable context
All three surfaces share one limit: they live inside ChatGPT and do not travel. Switch assistants and your custom instructions, memory, and projects stay behind. For context you want to own and reuse across tools, a dedicated memory layer like MemX holds it in a store you control and can search in plain language, kept private by architecture. The three ChatGPT features are still worth using well. They just are not where your long-term knowledge belongs.
01Is ChatGPT memory the same as custom instructions?
No. Custom instructions are a fixed field you set by hand that applies to all chats. Memory is updated automatically as you chat and reuses facts it decides to keep. One is manual and stable, the other is automatic and evolving.
02When should I use a Project instead of memory?
Use a project when work needs its own files and should stay isolated from your other chats, or when the topic is sensitive. Memory is for general facts you want carried everywhere, which is the opposite of isolation.
03Can custom instructions and memory conflict?
Yes. An auto-saved memory can contradict your custom instructions, and the model then reconciles them unpredictably. If behavior seems off, review memory first, since it changes on its own while custom instructions do not.
04Do projects have their own instructions?
Yes. A project can carry custom instructions and files that apply only inside it, which is what keeps project work separate from your general chats and context.
05Do memory and instructions transfer to other AI assistants?
No. Custom instructions, memory, and projects stay inside ChatGPT. To reuse context across assistants you need an external store that you control rather than relying on one platform's features.
