The context window is not the context

A technical note.

Every few months the context window gets bigger, and every few months someone declares that memory is solved. It is not. A bigger window is more room to paste things into. It is not knowledge of the user.

Two problems survive any window size.

The first is retrieval. If the model can hold a million tokens, something still has to decide which million. The hard part was never the holding. It was choosing what belongs.

The second is persistence. A window is per-session by construction. Close the chat and the context is gone, unless something outside the model wrote it down. That something is the real memory system, and it is usually missing.

So the window is a buffer, not a store. Treating it as a store is why assistants forget you between sessions and re-learn you every morning.

The useful question is not how big the window is. It is what writes to the store, what reads from it, and who owns it. That is a data problem, and it sits outside the model.