A million tokens. So why does your AI still forget?

A context window describes what a model can reference while generating a response. It is different from the material learned during training and from a product feature that stores information between conversations. Anthropic’s documentation counts system instructions, messages, tool results, images, documents and tool definitions toward the working window. New output and thinking also consume capacity.

This means a long agent session can fill its allowance without a user typing an enormous question. Screenshots, search results and earlier answers accumulate too. The documentation also warns that accuracy and recall can degrade as context grows. A larger limit makes more material admissible; it does not guarantee equal attention to every passage. Compaction can shorten older history so work continues, but the summary becomes the retained representation. The original document still needs to remain available somewhere if a later answer must be checked against its exact contents.

Google’s Gemini guidance similarly presents context as short-term working information and describes long documents, audio and video as possible inputs. It also shows why a model name matters when reading capacity claims. Gemini 4 shares its million-token window across input, thinking and output, whereas earlier Gemini generations have separate input and output budgets. Filling the input allowance can therefore leave less room for the answer under one accounting scheme than another.

Long context can reduce the need to discard material or retrieve only a few snippets, but it does not make selection irrelevant. Put the relevant question and evidence in a form the model can use, then inspect whether its answer actually cites the needed passages. For video, tokenisation and sampling also affect what enters the request. Before treating an assistant as a durable archive, identify where the source files live, which material is passed to each request and what the product retains after the session ends.