Set memory and custom instructions
Teach stable facts, not every passing preference
After this, you'll be able to separate memory, custom instructions, Project context, and one-time chat instructions without mixing them together.
Before you start
Complete Build your first Project first.
The idea
Memory is for durable facts, custom instructions are for standing preferences, Projects are for recurring context, and the current chat is for today's task. When those layers blur, ChatGPT starts carrying the wrong assumptions forward.
- GoalAfter this, you'll be able to separate memory, custom instructions, Project context,
- InputsMemory should contain facts that stay true across many chats
- Set memory and custom instructionsStable facts are separated from temporary project details
- ReviewCustom instructions describe response preferences, not source material
- ProofSort ten context items into memory, custom instructions, Project context, current
Here is the before and after: before, ChatGPT is answering a loose request. After, you can separate memory, custom instructions, Project context, and one-time chat instructions without mixing them together.
Now try it use the exercise prompt on one real task from today. Keep the answer small enough to check before you reuse it.
You are ready when the task, source, surface, and proof are all visible.
Try it (12 min)
Watch out for
- Saving temporary project details as account memory.
- Putting long source documents into custom instructions.
- Leaving wrong memories because they only fail in subtle ways.
- Using memory as a substitute for uploading the actual source file.
Paste this into Claude
Help me sort my ChatGPT context into the right layer. Facts and preferences: [paste 8-12 facts, such as role, company, writing style, current project, temporary deadline, recurring audience, preferred format, one-time task] Sort each item into: 1. Memory 2. Custom instructions 3. Project context 4. Current chat only 5. Do not save For each item, give one sentence explaining why. Then write the exact 5-7 sentence custom-instructions draft I should use.
What a good response looks like
Memory: You run post-production at Buck, because that role is durable context. Project context: June finance deck, because it belongs to a specific workstream. Current chat only: the draft title, because it may change today. Do not save: one-off tone experiment.
What good looks like
- Stable facts are separated from temporary project details
- Custom instructions describe response preferences, not source material
- Temporary details are not saved as memory
- The final instructions are short enough to review later
When this breaks
- Breaks when memory contains stale facts because future answers inherit old assumptions.
- Breaks when custom instructions carry task details because every unrelated chat starts with the wrong frame.
AI can help with this
Use ChatGPT to help you you can separate memory, custom instructions, Project context, and one-time chat instructions without mixing them together. Start with the exercise prompt and your real input. Ask for one draft, then check it against this proof: Stable facts are separated from temporary project details. Accept only the version you can verify yourself.
You can now
You can define what belongs in memory
- You can write short custom instructions
- You can keep Project context separate
- You can remove stale remembered facts
Key takeaways
Context has layers. Save durable facts, write stable preferences, keep project context inside Projects, and leave one-time details in the chat.
- Memory should contain facts that stay true across many chats.
- Custom instructions should shape answers, not store documents.
- Projects hold recurring work context.
- Wrong memory is worse than no memory until you correct it.
Go deeper