You've noticed it: ChatGPT starts strong, then after 10-15 messages, it forgets what you said earlier, contradicts itself, and gives generic answers. This isn't a bug — it's a feature of how AI works called the "context window." Understanding it takes 5 minutes and will permanently improve how you use every AI tool.
- What it is: The maximum amount of text an AI can "hold in mind" during a conversation
- ChatGPT (GPT-5.4): 256K tokens (~190,000 words)
- Claude (Opus 4.7): 1M tokens (~750,000 words) — largest available
- Gemini 2.5 Pro: 1M tokens
- Why conversations degrade: Earlier messages get compressed or dropped as the window fills
- Fix: Start fresh conversations for new topics
- Last verified: April 2026
How It Works (No Technical Jargon)
Imagine reading a book, but you can only keep 100 pages in your hands at once. As you turn to page 101, page 1 falls to the floor. You can still reference the recent pages clearly, but the early pages become fuzzy or disappear entirely.
That's a context window. Every message you send, every response the AI generates, every document you paste — it all counts against the window. Once the conversation history exceeds the window, the AI starts compressing or dropping earlier content. The result: it loses track of instructions you gave earlier, repeats things it already said, and gives less specific answers.
Why This Matters for Your Work
If you're working on a long document with Claude and you've been iterating for 30 messages, the AI might have lost your original instructions about tone, formatting, or audience. The fix is simple: start a fresh conversation and include your key instructions in the first message.
If you're using Claude Projects, the Project context (uploaded documents and instructions) stays persistent across conversations — but the conversation history within each chat still has limits. This is why Projects are so powerful: your core context never drops out.
Practical Rules
Start a new conversation for every new topic. Don't use one conversation thread for everything — each new topic gets a fresh context window.
For long projects, periodically summarize your progress and paste it into a new conversation. "Here's where we are: [summary]. Continue from here."
When conversations get weird, it's usually a context window issue. The fastest fix is always a fresh start with clear instructions.
Shorter conversations produce better results because the AI can fully attend to every message. Quality degrades with length, not time.
For a comparison of context window sizes across all major AI models, check our State of AI Models page.
📬 This is what we do every week.
One actionable AI insight per week. Plus a free prompt pack when you subscribe.
Subscribe free →Disclosure: Some links in this article are affiliate links. We only recommend tools we've personally tested and use regularly. See our full disclosure policy.