Why ChatGPT forgets what you said earlier (and how to fix it)

You told ChatGPT the rules three replies ago, and now it's ignoring them. This isn't a bug or a bad day — it's how the model reads a thread, and there's a clean way around it.

The short version

ChatGPT doesn't remember your conversation the way you do. Every time you send a message, the model re-reads the thread from the top through a fixed-size window. When a chat gets long enough, the oldest parts stop fitting — so the instruction you gave early on quietly falls out of view. The model isn't disobeying; it genuinely can't see that part anymore.

The fix is not to argue with it or repeat yourself louder. It's to start a fresh chat and carry forward only what matters. Here's how to spot the problem and do exactly that.

Why it happens: the window fills up

Think of the model's attention as a fixed budget of space. Your system setup, every message you've sent, and every reply it's given all compete for that same budget. Early in a chat there's plenty of room, so instructions stick. As the thread grows — long pasted documents, back-and-forth debugging, tangents — the budget fills. Something has to give, and it's usually the material furthest back: your opening instructions, the constraints you set, the details you mentioned once and assumed were locked in.

That's why the failure feels so specific. Recent turns are crisp; the rules from the start of the session are the first to blur. The longer the thread, the more of the beginning is out of reach.

How to spot it

You're almost certainly hitting this if you notice:

• It breaks a formatting or tone rule you set at the start, even though it followed it earlier.
• It re-asks for a detail you already gave — a name, a number, a preference.
• It contradicts a decision the two of you already made.
• Reminding it works for one reply, then it drifts again.
• Answers feel more generic and less anchored to your specific situation than they did an hour ago.

If that's your thread, no amount of rephrasing fixes it — the context is the problem, not the wording.

The fix: fresh chat plus a handoff

Reminders inside a bloated thread just add more text to an already-full window. The reliable move is to open a new chat and hand it a tight summary of what came before — your goal, the decisions that are locked, the constraints that keep getting dropped, and where you are right now. A fresh chat with a clean brief behaves like the model did in the first ten minutes: sharp and on-instruction.

1. Scroll your current chat and note the essentials: what you're trying to do, any decisions already made, the rules that keep getting forgotten, and the current state.
2. Open a brand-new ChatGPT chat — don't keep pushing the old one.
3. Paste a short handoff brief as your first message (template below).
4. Confirm it has the rules straight, then continue where you left off.

Copy-paste handoff template

Fill this in from your old thread and paste it into the new chat:

I'm continuing work from an earlier conversation. Here's the context you need.

GOAL: [what I'm trying to accomplish]

DECISIONS ALREADY MADE (do not reopen these):
- [decision 1]
- [decision 2]

RULES TO FOLLOW EVERY TIME:
- [format / tone / constraint 1]
- [constraint 2]

WHERE WE ARE NOW: [the current state / last thing done]

NEXT: [what I need you to do next]

Acknowledge these, then let's continue.

Keep it from happening again

Put your must-follow rules in the first message of any new chat, not scattered across replies. When a thread starts drifting, treat it as a signal to hand off early rather than fighting it for another twenty messages. And keep genuinely separate tasks in separate chats so one long tangent doesn't push out the rules for everything else.

Related: Why long AI chats get worse over time (context rot) · How to hand off an AI conversation to a fresh chat