The context window is the information available to a model during a particular interaction. It may include instructions, conversation history, documents, tool results, and space for the reply. Tokens are the units used to represent that content; a token is not always a whole word.
Imagine a desk. Context is what is laid out for the model right now. Saved memory is in a drawer: the app must bring it onto the desk before the model can use it.
Capacity is not a guarantee
A long context window gives you room for more material. It does not prove that the model will find every relevant detail or connect distant facts correctly. Google’s long-context guidance discusses both the uses and limitations of working with large inputs.
Why placement can matter
The 2023 research paper Lost in the Middle found that the tested language models could perform differently depending on where relevant information appeared. That result describes the models and tasks in the study; it is not a current ranking or a universal score for today’s models.
Give the material structure
For a practical experiment, label documents clearly, ask a specific question, and request a citation or passage supporting the answer. If the task depends on one small detail, verify it in the source. When the input is unwieldy, try a focused set of documents and compare the outcome.
Persistent memory is a different feature
A product may save preferences or retrieve old information and add it to a later interaction. That is separate from the amount of context available in the current request. “The app remembers me” does not mean every past conversation is visible to every model call.
Local models have a memory budget, too
Increasing context generally increases runtime memory use. That is why the finder lets you change conversation length alongside hardware and workload. A configuration that fits a short chat may become uncomfortable with long inputs.