Notes on coding agents and the pages they leave behind, on moving between notes apps, and on the plain mechanics of keeping what you think. Written and published from a Sumibako vault.
ChatGPT says "Context window full": what it means, what upgrading buys, and how to carry on
ChatGPT's Context window full message explained: why it appears after ten messages, what upgrading changes, and how to carry the chat on.
Prashant Shaw
Founder, Sumibako
6 min read
Since the end of October, some ChatGPT conversations stop with a new message: "Context window full: Start a new chat with an automatic summary of this conversation, or upgrade to continue here with more context available in every chat." If that sentence sounds familiar, it should. For a month before, the same wall said "Chat memory full", and I wrote a whole post explaining that it had nothing to do with memory. ChatGPT has now, apparently, read it.
This post is what the new message means, why it can arrive after ten messages, what pressing "upgrade" actually buys you, and how to carry on without losing the conversation. It was checked on 3 November 2026 against OpenAI's forum, where the reports and OpenAI's own replies are, and its pricing page. OpenAI has no help article for this message yet.
🫙 The same wall, with an honest name
Put the two messages side by side:
Late September to late October
Since about 30 October
Headline
Chat memory full
Context window full
What it offers
"Start a new chat with an automatic summary of this conversation, or upgrade to continue here with expanded memory in every chat."
A new chat with a summary, or an upgrade "with more context available in every chat"
Everything except the label and the last few words is identical, which is why I think it is the same limit with a better name, though OpenAI has not said so. The old label sent thousands of people into Settings to delete their saved memories, which never helped. OpenAI's support account said as much on the forum on 2 October: "The message indicates that the conversation reached its available memory/context limit." And again on 25 October: "The warning is about how much history ChatGPT can use in that chat. You don't need to delete your saved memories."
So: this is not about your saved memories, your custom instructions, or your account being full. It is about one conversation having used up the room ChatGPT gives it.
📏 What a context window is, and how big yours is
The context window is how much text a model can consider at once: every message in the chat, its replies, any files, plus some things you never see. OpenAI's pricing page lists the window per plan, and I copied the numbers into my guide to ChatGPT's limits:
Plan
Everyday answers (Instant)
Reasoning models
Free
27K tokens
Varies
Go and Plus
54K tokens
256K tokens
Pro
128K tokens
400K tokens
A token is roughly three quarters of an English word, so 27K tokens is a long short story. Part of it goes before your first message: system instructions, your saved memories if memory is on, and the tools the chat can use. In a conversation with long replies, pasted documents or a story you are writing together, 27K goes faster than you would think.
🐛 Why it can arrive after ten messages
The trouble is that people are not only hitting it in long conversations. In the forum thread that opened on 30 October, one person writes that in a fresh chat, "within 10 prompts (yes, I counted this time)", the notice was back. Another, writing fiction, gets "only 15-20 messages and boom!" They report it on iOS, the web, Windows and macOS. A third says "Continue in new chat" made a branch with no summary in it, which then hit the wall after one more message.
The most careful report I found is older and in Japanese: a Free user on Android who tested it in a temporary chat with personalisation off. Long messages, split messages and conversation alike, the warning arrived at about the tenth exchange. Sending one character at a time, "1", "2", "3", it arrived at 44. Forty-four characters do not fill 27,000 tokens, so whatever stops those chats is not only the size of the window.
OpenAI has been here before. In mid October the old message blocked short and brand-new chats for many people until OpenAI applied a fix on 21 October, which I covered at the time. On 25 October, after reports that it was back, OpenAI's support account wrote that it was "investigating reports" that the warning "stops people from sending messages after only a few exchanges." The new thread had no reply from OpenAI when I checked.
My reading, which is a reading and not a fact: a long conversation hitting this is working as designed, and a short one hitting it is a bug or a limit OpenAI has not described. Either way, the fix on your side is the same.
💳 What "upgrade" buys you
The message offers an upgrade, and on 2 October OpenAI's support account suggested the same: "If you need more capacity for longer conversations, you can also check the available plans to see which option offers higher limits."
Here is what that means in numbers. Going from Free to Go or Plus doubles the everyday window, 27K to 54K tokens. Going to Pro takes it to 128K. The reasoning models have much bigger windows, 256K on Go and Plus, but they also use more of your allowance, and from December that allowance is shared with Codex on paid plans.
An upgrade moves the wall further away. It does not remove it, and if what stopped your chat after ten messages is the bug, paying will not necessarily fix it: people in the forum say Plus subscribers see it too. I would not upgrade to escape this message. I would upgrade if you were already hitting other limits and this is the last straw.
🧯 What to do
1️⃣ If the chat is short, it is probably not you. Reload, try again later, and add the details OpenAI asks for to the forum thread: your plan, your platform and roughly when it happened.
2️⃣ Press "Start a new chat" if it is offered, then check what it carried. Ask the new chat to say in two sentences where things stand. If the answer is vague, the summary was thin, and you should paste the missing parts in yourself.
3️⃣ If the summary is missing altogether, as some people report, ask the old chat for a handoff note if it will still answer: what we are doing, what has been decided, what is in progress, what comes next. If it will not answer, copy the parts that matter by hand, or export the conversation so nothing is lost.
4️⃣ Keep the long-lived part outside the chat. Characters, decisions, the plan, the rules of your world: one document you paste or upload at the start of each new chat. One person in the forum calls theirs a "series bible", which is the best name for it I have heard.
5️⃣ Leave room. Shorter custom instructions and fewer saved memories leave more of a small window for the conversation itself, especially on Free.
Every version of this message teaches the same lesson: a conversation is a working surface, not storage. Anything you would be sad to lose belongs in a document you own.
That is where I come in, so weigh this accordingly. I make Sumibako, a notes app that AI apps can read and write pages in. On ChatGPT, OpenAI allows that through a custom connector on Business, Enterprise and Edu, and reading only on Pro; on Free, Go and Plus you paste, which works fine. Claude can read and write it on every plan, Free included. The page outlives the chat, whichever wall the chat hits.
ChatGPT has stopped telling you its memory is full and started telling you its context window is full, which is progress. Next year, perhaps, it will tell you how big it is.
