An AI context window is the working space available to a model during one request. It contains the material the model can consider while creating an answer. The service usually measures that space in tokens. Tokens are short pieces of text rather than exact words.
The window may hold more than the latest message. An assistant can receive its operating instructions, earlier chat turns, text from uploaded files, search results, and tool output. The response also needs room. A request that fills the whole limit leaves no space for a long answer.
Conversation memory has a boundary
Imagine a chat with a 20,000-token limit. The instructions use 1,000 tokens. Earlier messages use 12,000. A pasted document adds 5,000. Only 2,000 remain for the next reply. The application must trim something or ask for a shorter output.
This explains why an assistant may lose an old detail during a long session. The software might remove early turns once they no longer fit. Some apps summarize older discussion and place the summary inside the window. A saved profile or external memory can restore selected facts later. Neither method makes the model’s context unlimited.
More room does not mean perfect recall
A large context window lets a model accept long contracts, code files, or meeting transcripts. It does not guarantee equal attention to every line. A key fact can get lost among repeated or unrelated material. Clear labels and relevant excerpts often work better than a large unfiltered upload.
Models may also struggle to connect facts placed far apart. Tests of long-context performance use questions whose answers depend on details scattered through a document. Results can vary with document order, writing style, and the amount of distracting text.
How applications manage the space
A support assistant rarely needs every page in a company handbook. A retrieval system can search the handbook and insert a few matching passages. This method is called retrieval-augmented generation. It saves space and keeps the prompt focused.
Code tools may send the open file, nearby functions, and definitions that connect to the current change. Meeting tools may divide a long recording into sections before creating one final summary. These choices shape the answer as much as the published window size.
Context and training serve different roles
Training changes the model’s internal numbers over time. Context gives the existing model temporary material for the current task. Pasting a policy into a chat does not retrain the model. The policy can guide the present answer and then disappear from view on a later request.
Sensitive text inside a window still needs protection. Applications should limit access and avoid sending private data that the task does not need. A bigger window makes it easier to provide rich context. It also makes careless data sharing easier.