A chat has been running for days, and you refer to an instruction from its beginning. The response behaves as though the instruction was never given. The visible conversation and the material supplied to a model at one moment are not necessarily the same thing.
Understanding a context window helps explain that difference without treating the model like a person who simply forgot.
Think of context as the current working material
A model receives input within a finite capacity, commonly described in tokens. Tokens are units used to represent text and, in some systems, other media; they are not a universal one-to-one count of words. Limits and accounting vary by model and application.
Google’s long-context documentation explains the role of larger input windows, while its token documentation describes how its API counts and reports usage. Consult the current documentation for the system you use rather than transferring one model’s numbers to another.
A chat application may manage the history
An application can choose which messages, files or summaries to include. It may also offer a separate memory feature that stores selected information. A large visible chat therefore does not prove that every earlier sentence is supplied in full with each request.
Likewise, attaching a file does not by itself establish that every part is read on every turn. The application may retrieve selected sections or process the file through another mechanism. Check its documented behaviour when the distinction matters.
Keep essential instructions explicit
For an ongoing project, maintain a short, current task brief containing the objective, constraints and approved decisions. Remove superseded instructions instead of accumulating contradictory versions. Supply the relevant brief through the application’s supported workflow.
For example, a drafting task might need the intended audience, required language and a current terminology list. It rarely needs every abandoned opening paragraph from an earlier session. This is an organisation technique, not a guarantee of perfect model compliance.
Test the workflow you actually depend on
When evaluating an application, check whether important constraints survive realistic long tasks and file changes. A headline context capacity does not measure every aspect of retrieval, reasoning or instruction-following quality.
Record the model and application version where available, the material supplied and the result. Avoid placing confidential information into a service without appropriate authorisation. The practical lesson is to manage the working context deliberately, rather than assuming that a long conversation automatically becomes a dependable permanent memory.