Stage 2 · What happens when you write to it
Context window
What can the model "see"?
Everything the model has in front of it when it answers: system instructions, the conversation, attached files and tool results. It has a maximum size measured in tokens. If it isn't there, it doesn't exist for the model.
A context window halfway through a conversation (illustrative):
System instructions 8%Available tools 10%Conversation 34%Attached file 28%Free 20%
An image to remember it
The desk. Whatever isn't on the desk isn't at hand.
The misunderstanding
It remembers what we talked about last week. Only if the application puts that information back in the window. Memory is a feature of the harness, not of the model.
What changes in practice
New topic, new conversation. And give it what's relevant, not everything you have.
