The context window is a desk, not a memory

A model does not remember your last conversation. On each turn it reads everything sent to it: system instructions, chat history, pasted files and tool results. The limit on all of that combined is the context window.

What this means in practice

• Paste only what the task needs, not the whole folder.
• At the end of a long stretch, ask for a written summary of the decisions, then start fresh with it.
• Put rules you repeat into a project instructions file.

Still unclear to me

Tools compress long conversations differently, and they rarely say what survives. When a model “forgets” a decision, suspect compression first.