Why AI 'Forgets' Halfway Through a Long Chat

AI has a memory limit called the context window. Hitting it explains a LOT of weird behavior.

What to actually do

• GPT-5 class models hold roughly 200k–1M tokens (a novel)
• Older or free models are much smaller (8k–32k)
• When you hit the limit, the start of the chat literally disappears from memory

The big idea: AI has a memory limit. Knowing where it is keeps you from being surprised when it forgets.

AI models can only 'see' a fixed number of tokens at once — that's the context window. When chats get long, the AI literally forgets the top.

Some examples

• GPT-4o handles about 128K tokens — roughly a short novel.
• Claude can handle around 200K tokens.
• When you say 'as I mentioned earlier,' the AI may not remember.
• Long chats run out of memory; start a new one for new tasks.

Try it!

Open your longest ChatGPT thread and ask 'what was the very first thing I asked you in this chat?' — see if it gets it right.

AI has a limited context window — usually 100k to 1M tokens. Once you hit it, the model starts forgetting the early part of your chat. Knowing this changes how you structure long projects.

How to use it

• Ask AI for the context window of the model you're using
• Start a new chat for unrelated topics to save context
• Ask AI to summarize your chat so far to refresh memory
• Use Projects or Memory features to persist info across chats

Try it

In a long chat, ask AI to summarize your conversation so far. Then ask 'what was my first question' and check accuracy.

The context window is how much the AI can 'remember' in one conversation. In 2026 the spread between models is huge, and choosing wrong wastes hours.

Some examples

• 1M tokens ≈ 750,000 words ≈ 7 novels
• Use Claude 1M for: whole codebases, every chapter of one book
• Use GPT-4o 128K for: a long paper, a chat history
• Gemini 2.5 Pro: 2M tokens, even bigger but slower

Try it!

Pick your biggest text source (a long PDF). Try summarizing it in ChatGPT first. If it errors on length, retry in Claude. Note the difference.

Every LLM has a 'context window' — a maximum number of tokens (roughly: word-pieces) it can hold in its short-term memory at once. When the conversation runs longer than that, the oldest messages drop off and the AI 'forgets' them. Modern models range from 8K tokens (very small) to 1M+ tokens (Gemini, Claude Sonnet 4.5). Knowing your model's window prevents the classic bug: 'wait, I told you that an hour ago, why don't you remember?' For long projects, use a model with a big window or save context in a 'system message' or Project.

Some examples

• 1 token ≈ 0.75 words in English — so a 100K token window holds roughly 75,000 words, about a 250-page book.
• Free ChatGPT (GPT-4o-mini) is 128K tokens; Claude Sonnet 4.5 is 1M; Gemini 2.5 Pro is 2M. Long-doc tasks favor Claude or Gemini.
• ChatGPT's 'Memory' feature is separate — it stores facts about you across chats but doesn't expand the context window of any single chat.
• In Claude Projects or ChatGPT Custom GPTs, you can pin instructions and files that persist — a workaround for context-window limits.

Try it!

Try this in ChatGPT: paste a long article (5,000+ words), ask 5 questions about it, then ask 'what was the second paragraph about?' If it stumbles, you just hit the context-window edge.

Every AI model has a 'context window' — the amount of text it can hold in attention at once. When you go over, the model literally forgets the start of your conversation or document. Knowing the limit, and how to manage what fills it, separates power users from frustrated ones.

Some examples

• GPT-4 turbo: 128k tokens. Claude: up to 1M. Gemini: up to 2M. Your prompt eats from this budget.
• Long conversations gradually 'lose' the early instructions — restate them when needed.
• For long documents, summarize sections rather than pasting everything raw.
• Persistent 'memory' features add specific facts back into context every time.

Try it!

Look up the context window of the AI tool you use most. Plan your next big project around it.