What's the actual difference between Custom Instructions and Memory?

Custom Instructions is a static block of text you write once that gets prepended (in spirit) to every conversation — 'here's who I am, here's how I want you to respond.' Memory is dynamic: ChatGPT decides what to save as you talk, based on what it thinks is worth remembering. Custom Instructions you control fully; Memory you only review and edit after the fact.

Do they replace each other?

No — they layer. Custom Instructions are your stated, intentional context. Memory is the running observation of how you actually behave across conversations. The best setups use both: a tight Custom Instructions block for the things you always want true, and Memory to capture the patterns you didn't think to write down.

Which one should I lean on more?

Custom Instructions, if you want predictable behavior. Memory, if you want emergent behavior. For most people, Custom Instructions does 80% of the work and Memory adds polish. The mistake is dumping everything into Memory and hoping ChatGPT picks the right bits to use — it won't, consistently.

Why does ChatGPT sometimes remember things I never told it to?

That's Memory doing its thing. It scans conversations and saves entries it thinks are worth keeping. You can view all of them under Settings → Personalization → Manage Memory. If something landed there that you don't want — or that's wrong — delete it. The list is editable.

Do Custom Instructions and Memory survive model updates?

Custom Instructions yes, mostly — they're stored as plain text and re-injected each time. Memory survives the data, but the way ChatGPT uses Memory shifts with model updates. The same saved memory may be drawn on heavily by one model version and ignored by the next. That's why a portable persona you control is the durable layer underneath both.

Should I keep my Custom Instructions somewhere outside ChatGPT?

Yes. They're some of the most refined writing you've done about yourself — what you do, how you think, how you want help. Save them in a notes app or vault you control. If ChatGPT changes the field size, deprecates the feature, or you switch tools, the work survives.

A short explainer · ~4 min read

ChatGPT Custom Instructions vs Memory: what's the difference?

Two settings, two purposes, constantly conflated. One you write. The other ChatGPT writes about you. They layer in a way that matters — and they fail in different ways when OpenAI ships a new model.

Custom Instructions

What: A static text block you write once.

Who controls it: You.

When it applies: Every conversation.

Good for: The things you always want true — your role, your tone preferences, how you want answers structured.

Memory

What: Dynamic entries ChatGPT saves as you chat.

Who controls it: ChatGPT writes; you edit/delete.

When it applies: When the model decides it's relevant.

Good for: Things you didn't think to write down — recurring projects, relationships, ongoing context.

How to use them together

Custom Instructions gets the must-haves. Who you are, what you do, the format you like answers in, the things you never want ChatGPT to do. Write it tight; long isn't better.
Memory gets the rest. Let ChatGPT capture the running context — current projects, the names of people in your life, recent decisions — as it comes up naturally.
Review Memory monthly. Settings → Personalization → Manage Memory. Delete anything wrong or outdated. The model uses what's there; bad entries make answers worse.
Keep both backed up outside ChatGPT. Copy your Custom Instructions to a notes file. Export your conversation history (Settings → Data Controls → Export) monthly. Both are work you've done; neither lives anywhere you control.

What changes with model updates

Custom Instructions are durable — the text doesn't change, and OpenAI re-injects it on every model. Memory is more fragile. A new model may weight Memory entries differently, ignore some, lean too hard on others. You'll notice it as continuity that quietly shifted: ChatGPT remembering wrong details, forgetting recent ones, or surfacing old memories at strange moments. That's not your imagination. That's the model change.

Short answer

Custom Instructions = what you tell ChatGPT about you, once, deliberately. Memory = what ChatGPT decides to save as you talk. Use both: Custom Instructions for the must-haves, Memory for the rest. Back both up outside ChatGPT — they're work you've done that doesn't live anywhere you control.

The portable answer

Konshus is the vault both should live inside. Your Custom Instructions become a stable, versioned artifact you can paste into ChatGPT, Claude, Gemini — anywhere. Your conversations get imported and distilled into a portable persona that captures the same patterns ChatGPT's Memory is trying to track, but owned by you and usable across providers.

See also what to do when ChatGPT memory fills up and Konshus vs ChatGPT memory for the longer view.

Frequently Asked Questions

Both layers, owned by you.

Konshus turns your Custom Instructions and your conversation history into a portable persona that works in any AI — even after OpenAI changes how Memory behaves.

Meet Konshus