04 · The window
Context
The context window is working memory. If it is not in the window, the model cannot use it. Context engineering is the craft of what to put there.
In English
The model can only use what fits in its window — a limited working memory. System instructions, tools, retrieved docs, and the chat all compete for space. If it is not in the window, the model cannot “remember” it, no matter how important it was yesterday.
- Try this
- Turn system, tools, and docs on and off. Shrink the window. Watch older chat drop while the prefix tries to stay.
- Keep this
- If it is not in the window, it does not exist for this call.
Lab · pack a context window
Demo windows are tiny on purpose. Production models run 32k–1M.
- system25 tok
You are a precise assistant. Prefer tools over guesswork. If the window is full, say so.
- tools37 tok
tools: - web_search(q: string) - calculator(expr: string) - read_file(path: string)
- retrieved35 tok
[retrieved] Nimbus-4 payload is 12 kg. ISO 10218-1 certified. Dual-channel e-stop required near humans.
- user18 tok
Can Nimbus-4 work next to people in a freezer aisle?
- assistant18 tok
I would check the safety spec and payload. Looking that up.
- user20 tok
Also: who is the CEO, and when did the Series B close?
Weights are long-term memory. The context window is working memory. System prompt, tool schemas, retrieved documents, the conversation, the scratchpad — they all compete for the same budget.
Context engineering is the name for packing that budget on purpose: what to include, what to retrieve, what to compact when the window fills. Prompt engineering was a slogan. This is the actual job.