Context Windows

Context Windows

A context window is the maximum amount of text a language model can process at one time. Think of it as the model's working memory — everything it can "see" while generating a response.

Everything in the context window is available to the model: your message, the conversation history, any documents you've pasted in, the system prompt. Anything outside the window is invisible.

How Context Windows Have Grown

Early GPT models had a 4,096 token context window — enough for a few pages of text. Context windows have grown enormously since then.

As of October 2026: frontier models from the major labs now offer context windows of around 1 million tokens as standard, not as a special case.

200,000 tokens is roughly 150,000 words — about 500 pages of text. A 1M token context can hold entire codebases or long books.

Larger context windows mean you can have longer conversations, analyse longer documents, and provide more context — all of which tend to improve output quality.

What Happens When You Hit the Limit

When a conversation exceeds the context window, something has to give. Historically that often meant truncation: older messages silently dropped, with the model losing access to early parts of the conversation.

Many applications now handle this more gracefully instead: they compact or summarise the older parts of a long conversation before it overflows, preserving the gist rather than either keeping everything or silently discarding it. Exactly what gets kept, summarised, or dropped is still up to the application, not the model.

This is why very long conversations sometimes feel like the model has "forgotten" detail from earlier on, even when the app is summarising rather than dropping messages outright — a summary keeps the gist but loses detail.

Practical implication: for long tasks, be aware that context is finite. Re-provide key information if you think it may have scrolled out of the window or been reduced to a summary.

The Difference Between Context and Memory

This is a common source of confusion:

Context window = what the model can see right now

By default, language models have no persistent memory. Each conversation starts fresh. When you close the chat and start a new one, the model has no recollection of previous sessions.

Some applications build a memory layer on top of this — both ChatGPT and Claude now have memory features that can recall facts from earlier conversations. But this is done by the application, not the model itself. Behind the scenes, remembered facts are retrieved and injected back into the context window at the start of a new conversation.

True long-term memory doesn't exist natively in language models — it's always being simulated by clever context management.