context window

in English: context window

basicwidespread

The context window is the amount of text an AI model can see and consider at once when answering you.

In plain terms

Imagine working at a desk that can only hold a limited number of sheets of paper. What you place on the desk, you can see and use; what does not fit remains in the drawer, and you cannot look into it while working. The context window is the language model's desk.

Everything the model is currently working on goes into this window: your question, the earlier part of the conversation, uploaded documents, background instructions, and the answer being prepared. The model does not remember like a human; it re-reads everything in the window with each response.

Its size is measured not in words or pages but in tokens. A token is a word fragment: a Hungarian word typically consists of multiple tokens, so the same text takes up more space in Hungarian than in English.

This is not the same as the model's knowledge. The model acquired knowledge during training, and it does not change afterward. The context window is the current workspace: it restarts with every new conversation, containing only background instructions and data loaded from memory, if any.

How it works

The system breaks the entered text into tokens and places them sequentially in the window along with previous messages and attached materials.
The model considers the entire content of the window for each new piece (token) of the response and calculates what should follow from it.
Its own response also enters the window, so the conversation takes up more space with every turn.
When the window fills up, something must be removed: the application either truncates or compresses and summarizes the oldest parts, or signals that you need to start a new conversation.
Long-running agents therefore manage space separately: they write notes to a file, summarize work done so far, and load only what is needed for the current step.

An everyday example

You spend a long evening planning a vacation with a chatbot, and at the start of the conversation you wrote that your partner doesn't eat fish. Two hours later, the model recommends a fish dinner. It wasn't careless: your early message fell out of the context window, or was lost in a compressed summary.

Who is affected

everyoneThis explains why a chatbot 'forgets' during long conversations and why it is advisable to start a new conversation for a new topic.

employeeThis determines whether a model can process a long contract, report, or spreadsheet as a whole, or if you must split it. → Work & HR

developerA coding agent only sees the portion of code that fits in its context window, so how that space is allocated directly affects the quality and cost of the work. → Code

entrepreneur, executiveProviders bill per token, so a window full of text means higher costs and slower performance.

A common misconception

MythThe larger the context window, the better the model, and anything once included is remembered with certainty.

In realityA large context window means more space, not better attention: with very long texts, the model tends to overlook certain details, especially in the middle. Often, less but well-selected material is more valuable.

Limits and risks

What falls outside the window is unknown to the model, and it usually does not warn you about this but continues answering confidently.
The more text you input, the slower and more expensive the response becomes, and the easier it is for the core message to get lost in the details.
The window content is lost at the end of the conversation: the model will only know about it next time if the application has a dedicated memory function.
The advertised maximum size is not always available: the actually usable limit may be smaller depending on the subscription and application.

Similar, but not the same

training data – The model's permanent knowledge was built from the training data during training, while the context window is only the temporary workspace for the current conversation.
memory (in an AI assistant) – The memory function stores data about you in a separate repository between conversations and loads only as much of it into the window as is currently needed.
prompt – A prompt is what you type; the context window is the full frame that holds the prompt, history, attachments, and the response.

Related concepts

Broader

large language model, generative AI, transformer

Narrower

system prompt, context compression, output token limit

On the edge

token, prompt, context engineering, retrieval-augmented generation, hallucination

A brief history

The term became dominant with language models based on the transformer architecture introduced in 2017. These process text segments of up to a certain length simultaneously. The window of GPT-3 in 2020 was only around two thousand tokens. In 2024, with Google's Gemini 1.5 Pro model, the first widely available models with one million tokens appeared.

Where things stand

At the end of 2026, the context window of leading models typically ranges from several hundred thousand to one million tokens, and some models go beyond that. This is the equivalent of several books of text. The focus has shifted from sheer size to how to fill the window intelligently, especially for agents working independently for long periods.

Status: 2026 November

In the news

Appears in 2 Agent Age stories; the curve shows weekly occurrence in the news magazine.

Nobody knows who built free Ox Alpha coding model or where submitted code goesAug 22, 2026 · Code
SpaceXAI launches Grok Bot agent and releases Grok 4.6 modelAug 11, 2026 · Agent Age

Also known as

kontextusablak · kontextus ablak · kontextus-ablak · context window · kontextushossz · context length · context size

Sources

This entry was machine-edited and has undergone machine expert review; the “In the news” section updates automatically from Agent Age stories.