In short
Everything the model can see on one run: the instructions, the files it was given, the conversation so far. Outside the window, it does not exist.
A model has two kinds of memory. What it learned in training, which is vast and frozen. And the context window, which is what you put in front of it right now: the instruction file, the documents, the messages. Only the second one is yours to control.
This is why instructions live in files that get loaded every run, and why documentation matters. An agent does not remember yesterday unless yesterday was written down and put back in the window today.
It is also why instructions are short and the hard rules sit at the top. A window is finite, and attention inside it is not even.
Example
What it looks like.
context-window.mdplain text
At the start of a run the agent loads: its instruction file, today's and yesterday's notes, the card it was assigned, and the skill for that job. That is its whole world for the next ten minutes.Related
Read next.
- modelThe intelligence an agent calls when it needs to read, write, or decide. Replaceable. When a better one ships, the same instructions run on it and the results improve.Read the definition
- tokenThe unit a model reads and writes in. Roughly three quarters of a word. Models are priced and limited by tokens, not pages.Read the definition
- documentationThe written record of how a system works. Here, the agent keeps it current, so the person who was not in the room can take over.Read the definition
- instruction fileThe Markdown file that tells an agent who it is, what it owns, and what it may never do. The agent's job description, readable by the agent.Read the definition