You’re twenty messages deep in a chat, working through something real. Early on, you told the chatbot your budget is tight. Ten minutes later it suggests the expensive option — like you never said a word.
Sound familiar? That’s not the machine being rude. It ran out of suitcase.
Your chatbot has a suitcase
Here’s the one idea that makes everything else click.
A context window is the AI’s short-term working memory: the total amount of text it can “see” at once while generating one answer. Everything is measured in tokens — chunks of text, roughly three-quarters of a word each.
Now picture one suitcase. Your prompts go in. Any files you upload go in. The entire conversation so far goes in. And here’s the part nobody tells you: the answer the AI is about to write has to fit in the same suitcase too. It’s one shared budget, not free space for your documents. That’s why a model can cut off mid-sentence on a huge prompt — its own answer ate the room it needed.
So basically, the AI doesn’t remember you the way a friend remembers you. A friend has a lifetime. The chatbot has a suitcase — and closing the chat empties it. That is not the same thing as the knowledge baked in during training. Training is what the model learned. Context is what it can currently see.
Start with this if: you keep hitting weird forgetfulness in long chats. The problem isn’t your memory. It’s the suitcase being full.
— Roxie: Your own answers count against you. Every word the chatbot wrote back is sitting in the same suitcase, crowding out what you actually need it to keep. Long, chatty conversations fill the window faster than people expect.
How big is the suitcase?
Okay, the numbers. October 2026, straight from the companies’ own published plan pages — with the honest gaps marked.
| Plan | Price | Context window |
|---|---|---|
| ChatGPT Free | $0 | 27K tokens (Instant answers) |
| ChatGPT Go | $8/mo | 54K (Instant) · 256K (reasoning) |
| ChatGPT Plus | $20/mo | 54K (Instant) · 256K (reasoning) |
| ChatGPT Pro | $100/mo | 128K (Instant) · 400K (reasoning) |
| Claude (paid plans) | $20/mo+ | up to 1M tokens |
| Google AI Pro | $19.99/mo | 1M tokens |
| Gemini Free | $0 | 32K tokens |
OpenAI publishes its numbers right on its pricing page, split by mode: quick answers get the smaller “Instant” number, and the reasoning model gets the bigger one. One caveat the fine print hides: that’s the plan total — the slice actually available for your input is smaller.
Claude’s paid tiers get up to 1M tokens, per Claude’s pricing page. The honest gap: nobody publishes a per-chat window number for Claude’s free tier, so I’m not inventing one.
Google’s own help page gives Gemini 32K tokens free and 1M on Google AI Pro — with Google’s own math: 1M tokens covers roughly 1,500 pages of text, about 750,000 words. A few full novels, sitting in one chat.
Two deliberate omissions. Grok isn’t in the table because published numbers contradict each other — 128K to 2M depending on your model version — so any figure I printed would be a guess. Copilot isn’t here because Microsoft doesn’t publish a per-tier consumer number at all. That silence is the data point.
And one change worth knowing: as of October 9, 2026, free Gemini is Flash-Lite only. So free Gemini is now the smallest window and the weakest model of the big four’s free tiers.
Why a bigger suitcase isn’t a better memory
Here’s where it gets interesting. The obvious fix — just make the suitcase bigger — has a problem. Research says it doesn’t work the way you’d hope.
In 2023, Stanford and UC Berkeley researchers ran a study called “Lost in the Middle.” They gave models a question plus ten to thirty documents, one containing the answer. Accuracy was U-shaped: best when the answer sat at the start of the context, second-best at the end, worst in the middle. With 20 documents, one model scored about 76% when the answer was first — and about 54% when it was buried in the middle, worse than answering with no documents at all. And extended-context versions performed almost identically to shorter ones. A bigger window did not fix it.
Actually, think about what that means: you can stuff a million tokens of documents into a chat and the model still misses the fact buried deep inside. Google’s own help page admits it — answers over long input “may not account for all the content or may miss connections or details throughout the content.”
And this connects straight to the piece on hallucinations I just wrote: an overstuffed window doesn’t just get ignored — it can produce confident wrong answers. Same failure, different symptom.
— Roxie: Bigger windows are great marketing and mediocre memory. The industry sells the million-token number like it’s a library. It behaves more like a very long hallway where the model can only clearly see the doors at either end.
What hitting the limit feels like
There’s no gauge, no token count, no warning — ChatGPT shows nothing as the window fills. Here’s how it shows up:
1. It forgets what you set up early. This is the first sign, every time. The chatbot contradicts something you established earlier — the budget, the constraint, the name of the thing. The early messages got pushed out of the suitcase.
2. It starts summarizing instead of recalling. On Claude’s paid plans, when a long chat nears the limit, Claude summarizes earlier messages so the conversation can continue. The catch: the summary is lossy. Fine detail disappears — the specifics you cared about, gone.
3. Vague answers on long documents. You upload a big document and ask a question. The answer comes back smooth, general, and weirdly non-specific — it caught the beginning and the end and missed the middle where your answer was hiding. Classic lost-in-the-middle.
4. There’s no visible ceiling, so you blame yourself. You assume you explained it badly, or the tool is dumb today. No — the window filled. The model literally cannot see the material anymore.
The fixes that actually work
Good news: this is a workflow problem, not a “buy the bigger tier” problem.
Start a fresh chat when you switch topics. New topic, new chat — empty suitcase. This is the single highest-leverage habit in this whole article.
Summarize-and-carry-over. Ask the bot to summarize the key points and facts so far, then paste that summary into the new chat. You control what survives.
Upload only what you need. A 200-page PDF when you need five pages is paying for noise — and inviting lost-in-the-middle misses. Less context, more accuracy.
Use Projects for reference material. ChatGPT Projects, Claude Projects, Gemini Gems — filing cabinets, not memory. Files stay with that workspace, so you stop re-uploading the same documents into every chat.
Make NotebookLM your source-first move. When you need answers grounded in your documents, NotebookLM is the honest choice: upload your sources, ask questions, and get answers with inline citations to the exact passage. Google says it never trains its models on your uploads. For research-with-receipts, this is the tool that behaves the way people assume all AI behaves.
Know what “memory” actually is — and what it costs. Every assistant now carries things across chats: ChatGPT’s saved memories, Claude’s memory topics, Gemini’s Saved Info. The tradeoff is plain: memory lives on the provider’s servers, and every provider’s memory works only inside its own product — five walled gardens that don’t talk to each other. Turn it on for stuff you re-explain constantly. Use Temporary or Incognito chats for anything you’d shred on paper.
— Roxie: Memory features are convenience with a filing system. Every “remember this” is a line in a company’s database about you. Review what’s stored occasionally, delete what you wouldn’t put on a postcard — and know that turning memory off doesn’t always delete what’s already there.
The cost question: do you pay for a bigger window?
Here’s the part that surprises people: consumer subscriptions are flat-rate. The window size is a spec baked into the tier — nothing on your bill changes with how much context you use. Going from Free to Plus is the cost of a bigger window.
And the honest math: nobody fills 1M tokens in casual chat. The real question is whether your use case needs the bigger window — long documents, book-length projects, big codebases. If it does, the $20 tier is where the value is: Claude Pro and Google AI Pro both give you 1M tokens at twenty a month, while ChatGPT’s biggest consumer window lives on the Pro tier at $100/mo — five times the price.
One more ceiling check: message caps, upload limits, and deep-research quotas are separate walls, and they still bind. A 1M window with a throttled message cap still stops you. It just stops you differently.
Start with the $20 tier if: you work with long documents or book-length material regularly. Stay free if: your chats are normal-length conversations. The free windows are bigger than they look — 27K tokens is a long article, not a tweet.
The verdict: the one rule
The whole article compresses to one sentence: keep the suitcase light and put the important stuff on top.
Fresh chats per topic. Summarize before you switch. Upload the pages you need, not the whole library. Put the key facts at the start of your prompt — beginnings get found, middles get missed. And when you need answers grounded in real documents, use a source-first tool instead of stuffing everything into one chat.
The context window is the reason your chatbot “forgets.” It was never remembering. It was just reading from a suitcase — and now you know how to pack it.
Keep learning: Why AI Makes Things Up (And How to Catch It Before It Costs You) · AI Agents Explained: What They Actually Do (And the 5 Worth Trying) · How to Prompt AI Tools So They Actually Give You What You Want