Lexicon · II · Access
Context window
Also: context
Everything the model has in front of it when it answers: your request, the documents provided, the history of the exchange. Whatever is not in it does not exist for the model.
The analogy
The desk of a brilliant but amnesiac expert. He reasons remarkably well, provided everything he needs is on his desk at that very moment. File a folder away in the cabinet and he no longer knows it exists. Enlarging the context window enlarges the desk; it does not cure the amnesia.
The amnesiac expert's desk
Put documents on the desk and watch the gauge: whatever is not on it does not exist for him.
The documents
Check a document to put it on the desk, uncheck to remove it.
34,200 / 200,000 tokens
Cost and latency of every answer : low
Enlarging the window enlarges the desk. It does not cure the amnesia.
The orders of magnitude
A page of text weighs between 400 and 800 tokens depending on its density; a dense forty-page contract, about 30,000; recent models' windows range from 200,000 tokens to a million, sometimes more. Those figures sound immense, until you try to fit in an entire document base, six months of history or a system's logs. And the window is not a free cupboard: every token on the desk is reread with every answer, and paid for with every answer. Stuffing it "just in case" degrades price and speed at once, and often quality too, because the essential drowns in the rest. One last item, often forgotten: before your question even arrives, the harness lays its own pieces on the desk. Standing instructions, tool definitions, product rules: several thousand tokens, present in every request and billed with every answer.
A concrete example
Paste three contracts into the conversation and ask away: everything fits in the window, the answers are precise. Move to three hundred contracts: someone must choose what to put on the desk for each question; that is the harness's job (search, select, inject), and the real cause of most "it forgot what I told it ten minutes ago".
The memory tricks
When an AI product "remembers", look for the mechanism: there always is one, and it lives in the harness. Profile notes re-injected at the start of every conversation; running summaries that compress the history as it grows; a document search that selects, for each question, the relevant excerpts from a base too large for the desk (the famous "RAG"). Three tricks, one logic: choosing what goes on the desk. The organizational consequence is direct: answer quality depends on the quality of that choice, hence on your documents, their filing, their freshness. AI puts document governance back at the center of the game.
What it is NOT
It is not a document store: nothing persists in it. It is not infinite: it is counted in tokens, and it overflows. And it is not free: the fuller the window, the more each answer costs and the longer it takes; stuffing it "just in case" is an anti-reflex.
Often confused with
The distinction that changes the conversation.
- Memory
- The window empties when the conversation ends: the model retains nothing from one session to the next. When a product 'remembers' you, it is the harness re-injecting notes into the window at every exchange: an orchestrated stand-in for memory, not a memory of the model.