Working Memory
The active context the model can see right now — the desk. Every token in the prompt is active state, and it runs on a budget.
Working memory is the active context the model can see right now: the current request, the visible conversation, the instructions and files still inside the window, the tool results that just came back. In cognitive science, working memory is the limited mental workspace for the task in front of you — and the useful part of that analogy is the word limited.
Every token is active state
In agent systems, the context window plays this role. Every token in the prompt is part of the model's active state for this turn. If material is present, the model can reason over it. If it's absent, the model has to infer, search — or guess.
This is also the most basic memory experience people have with chat systems: mention the app root early in a conversation, ask about it later, and the model "remembers" — because the earlier message is still sitting in the same window. It appears to remember. Nothing was actually stored.
A bigger desk is still just a desk
When people say a model has a 200k or 1M token window, they mostly mean a bigger working-memory budget. That helps — a bigger desk lets you lay out more material. It does not give the system any way to remember after the conversation ends, and it doesn't remove the design problem:
Context window — a budget, not a bucket
FOCUSED
The window holds what matters. Attention is cheap.
Two ordinary ways it breaks
- It ends. Close the session and there's no durable state behind it. Open a new chat, say "use the same rules as before" — and the model has no reliable idea what before means. It may ask; it may guess. Even a good guess is still a guess.
- It fills up. The window is a budget. Instructions, history, retrieved snippets, tool output — every useful thing competes with every other useful thing.
So the first memory question is deceptively simple: what should be in front of the model right now? Working memory handles the turn in front of you. To cross the session boundary, the agent needs saved experience.