A context window is all the text a language model can take into account while writing a response. It includes earlier messages in the conversation, material you attach, your new , and even the response the model is generating. It is the model’s working memory, not the same as its training data.
What happens in a long conversation
Each message and response adds to what must fit in the window. The window is measured in , and its size varies by model and service. As the limit approaches, a service may drop the oldest messages or summarize the beginning of the conversation. A condition you gave at the start may then stop influencing the answers.
A larger window does not automatically make an answer better. When there is a lot of text, the model may pay less attention to individual details.
When a conversation gets long
- Provide only as much material as the task needs.
- Repeat a critical condition in a new prompt if you gave it far earlier in the conversation.
- Gather the key information into a short summary and, if needed, continue in a new conversation using that summary.