Why ChatGPT Forgets What You Told It (and How to Stop Repeating Yourself)
Short answer: two separate things are happening, and only one of them looks like forgetting. Inside one long conversation, the assistant is working from a window of text that has a limit, so instructions you gave near the top quietly stop carrying the weight they did an hour ago. Across conversations, nothing carries over at all unless you deliberately put it somewhere that gets read again. Neither problem is solved by a cleverer prompt. Both are solved by keeping your context outside the chat, where you control it.
What follows is the habit that replaces re-explaining yourself every morning: a short brief you keep in a note, a place to park it in ChatGPT, Claude or Gemini, and one move to make when a conversation has clearly drifted.
Two kinds of forgetting, and they need different fixes
The first kind happens inside a single chat. You set a rule early — keep it under 200 words, never use the word "artisanal", always address the reader as "you" — and for a while it holds. Thirty messages later the rule is gone. The assistant does not announce that it has dropped it. It simply produces something that breaks it, with the same confidence as before.
You can watch this happen rather than take my word for it. In a long chat, type: List every instruction I have given you in this conversation, in my words, in the order I gave them. The gaps in that list are the instructions that are no longer really in play. Anything the assistant paraphrases loosely is an instruction it is holding loosely.
The second kind happens between chats. A new conversation starts blank. The memory features that the big assistants now ship do carry something across, but what they carry is a short note the tool wrote about you, not a transcript of what you said. That distinction matters more than it sounds: you get a paraphrase, chosen by the tool, of what it judged worth keeping. It is a convenience, not a filing cabinet.
The brief that saves the most repeating
Almost everything you find yourself retyping falls into five buckets. Write them down once, in a plain note on your own computer, and you own the context instead of hoping the tool kept it.
What goes in your brief
- Who you are and what you do, in one line.
- Who the output is for — the actual reader, not "customers".
- What a good result looks like, described so someone else could judge it.
- Hard rules: length, format, words you never use, things you never claim.
- The mistakes it keeps making, written as instructions rather than complaints.
Filled in, for an invented example, it looks like this — short enough to paste at the top of any new chat:
Context for this chat. I run a four-person bakery and write our Friday customer email myself. The readers are regulars who already know us, mostly on phones. A good email is under 200 words, one thing to announce, warm but not cute. Never use the words "artisanal", "nestled" or "delight". Never invent an ingredient or a price. You keep opening with a rhetorical question — do not do that. Confirm you have this, then wait for my draft.
The last line matters. Asking it to confirm and wait stops it running off and writing the email before it knows what the email is about. If the tone half of that brief is what you struggle with, there is a longer method for it in getting ChatGPT to write in your own voice, and a general structure for requests in how to write better AI prompts.
Where to park the brief so you stop pasting it
All three of the main assistants have somewhere to put standing context. They behave differently enough that it is worth knowing which one you are relying on.
| Tool | Carried over on its own | Where you put something on purpose |
|---|---|---|
| ChatGPT | Memory notes it writes itself, when the feature is switched on | Custom instructions in settings; a Project, which keeps its own instructions and files |
| Claude | Less, by default — expect a new chat to start blank | Personal preferences in settings; a Project with custom instructions and reference documents |
| Gemini | Saved information, and it can draw on past chats where you allow it | Saved info in settings; a Gem, which is a reusable assistant with your instructions baked in |
The pattern is the same everywhere: the automatic thing is a convenience you did not write and cannot see the wording of, and the deliberate thing is a box you fill in. Put the brief in the box. Menus move — if you cannot find it where I have described it, open settings and search for "memory", "instructions" or "personalization" rather than assuming it is gone.
One practical note on Projects and Gems: they are per-topic, which is what you want. A bakery-newsletter Project that only ever sees bakery instructions will not start applying your bakery tone to a job application.
When a chat has drifted, hand it off instead of scrolling
The instinct when a long conversation goes wrong is to scroll up, find the instruction it dropped, and paste it again. That works once. By the third time you are managing a conversation instead of doing the work.
The better move is to close the chat deliberately and carry the decisions out yourself:
We are stopping here. Write a handover note for a fresh conversation. Include: the goal, every decision we have made, every constraint I gave you in my own words, and what is still unresolved. Do not include your own suggestions unless I accepted them. No advice, no summary of how well it went.
What the handover actually came back with
Two things worth noticing. It did quote the rules back accurately — the banned words and the no-rhetorical-question instruction survived word for word, which is exactly what you want from a handover. But it also moved one of my constraints, the line about readers being regulars reading on their phones, out of the constraints section and into a "Goal" paragraph it wrote itself. Nothing was lost here, and on a longer brief that is precisely how a rule quietly becomes background. Read the handover before you trust it; the quotation marks are the assistant's choice, not a guarantee.
Then open a new conversation and paste your brief and the corrected handover. You are starting from a clean window with the good part of the old one intact. This is also the right moment to save that handover into a Project or a note — it is the cheapest form of memory there is, and it is written in your words rather than the tool's.
Find out what it thinks it knows about you
If memory is switched on, it is worth looking at what has accumulated. There are two ways to check, and they do not always agree.
Ask directly — What do you have saved about me? List each item separately. — and then open the settings screen where the saved items are listed and read the actual entries. The list in settings is the real one. Delete anything stale, anything left over from a one-off task that has no business influencing future chats, and anything you would rather not have written down about your work.
This is a good habit for a second reason. Everything in that list came out of something you typed, which makes it a fair audit of what you have been feeding the tool. If that thought is uncomfortable, the boundaries worth drawing are covered in whether it is safe to paste work documents into ChatGPT.
The parts that bite people
Memory is a summary, not a record. Never rely on it for a number, a date, a name or a quote. If something matters, keep it in a document and paste it in. Anything an assistant recalls from an earlier chat deserves the same treatment as any other AI answer — the two-minute version of that check is in how to fact-check an AI answer.
Stale facts travel. A detail that was true in March — a job, a client, a plan you abandoned — can quietly shape an answer in September, and nothing flags that it is working from something old. When a project ends, prune the entry.
Saying "remember this" is a request, not a switch. Sometimes it saves; sometimes it agrees warmly and saves nothing. If it matters, put it in the box yourself rather than trusting the conversational version.
Work and personal context blur. The same memory that knows your writing style also knows whatever you mentioned about your employer. On a shared or family account, assume anything saved is visible to whoever else signs in.
And switching memory off entirely is a perfectly reasonable choice. The brief-in-a-note habit works without it, and it is the half of this you actually control. If you are still deciding which assistant to settle on, the everyday writing comparison covers how the three differ on the work itself.
If you change one habit
Keep a five-line brief in a note and paste it at the top of new chats. That single move removes most of the re-explaining, and it keeps working when a memory feature changes, or gets switched off, or writes down something about you that you never would have.
Related reading
- How to Get ChatGPT to Write in Your Own Voice — the longer method for the tone half of your brief.
- How to Write Better AI Prompts — a simple structure for the requests themselves.
- ChatGPT vs Claude vs Gemini for Everyday Writing — how the three compare on the actual work.
Written by Mitch, a software analyst who tests software for a living. Every guide here comes from actually using the tool on a real task — including the parts where it falls over. Tested on ChatGPT Plus, Claude Max and Gemini (free). More about this site · Corrections: [email protected].
Get the next guide by email
A new tested how-to when there is one — usually a couple a month. No hype, unsubscribe any time.
Tested in real accounts. No affiliate links.
