A context window is working memory
When an AI stops following an instruction from the beginning of a long chat, the conversation has not necessarily been deleted. That instruction may no longer be part of the information sent to the model for its next answer. The context window is the maximum amount of information the model can consider in one request. Messages, files and instructions occupy that space.
Google compares context to short-term memory. A desk with limited space is another useful picture. Your current question, earlier messages, attached documents and system rules compete for room. The model can work directly only with the material that the product puts on that desk for the current turn.
Context is measured in tokens, the units a model uses to split text and code. A token might be a word, part of a word or punctuation, so there is no exact word conversion. The OpenAI guide to tokens describes that process, while our explainer shows what AI tokens are and why they matter. Depending on the product, the window may also need to leave room for reasoning and the answer.
What happens when the chat no longer fits
A larger window can hold longer chats, books or software repositories, but it is still working memory for one request rather than unlimited human memory. Sending more input also tends to increase cost and response time.
Products handle overflow in different ways. They may remove old messages, summarise them, retrieve selected passages or ask you to start a new chat. An early instruction can therefore stop shaping the answer while the interface continues to display it. Stored history does not guarantee that the complete conversation reaches the model on every turn.
A model can miss a fact before the window is full. Very large contexts contain distractions, contradictions and outdated versions. The Lost in the Middle study found that several models retrieved information in the centre less reliably than material near the beginning or the end. Systems have improved since then, but maximum capacity still does not guarantee equal attention to every passage.
Our guide to how ChatGPT works completes the mechanism. The model produces an answer from the information supplied for that turn. It does not consult a flawless recording of every previous conversation.
History, memory and context do different jobs
History is the conversation that an application stores and lets you reopen. The context window contains the information included in a particular request. Memory is an extra feature that may preserve selected details between chats. A product can display a long history, use a smaller context window and have memory enabled at the same time.
Context is not the model's permanent knowledge either. Attaching a PDF lets the system answer from that file while it remains available, but it does not retrain the model. Search or RAG systems find relevant passages and put them into the context when they are needed.
How to stop AI losing important details
Begin each long phase with an operational summary that states the goal, settled decisions, facts the model must not invent and the required output. Put the current request at the end. If the direction changes, identify the earlier instruction that no longer applies.
Break large documents into tasks. You can ask for an outline first and then work through one section at a time. Keep the authoritative version in your own file, table or list and attach it again when accuracy matters. At the end of a session, request a short status note with open decisions that you can paste into the next chat.
When an answer contradicts an earlier decision, restate the correct fact instead of asking only whether the AI remembers it. Ask the system to identify the passage in its current context that supports the answer. That check helps separate a missing detail from an ambiguous instruction.
A bigger window helps but does not solve everything
Long windows make it possible to compare more files, follow large codebases and sustain complex conversations. OpenAI introduced GPT-4.1 with up to one million tokens of context in the API, and Google documents similarly large windows for some Gemini models.
The maximum figure does not reveal how reliably the model finds a detail, how much the request costs, how much room remains for the answer or how the product summarises older messages. Compare tools with a task that resembles your own and check whether important instructions survive.
Treat context as a large desk rather than endless storage. Keep the current objective and decisions visible, remove outdated versions and save a short summary before ending the session. For RAG, embeddings, multimodal models and agents, open our AI dictionary.
The conversation starts here
Sign in with a supporter account to comment. Sign in
Nobody has commented yet. Want to go first?