Context window lab

A model only sees what fits in its context window. Shrink the window and watch the oldest messages drop out of view.

What to notice

• The model takes the newest messages first. The oldest are the first to fall out.
• Once your opening rule is out of view, the model cannot follow it, however clearly you wrote it.
• Restating the rule brings it back into view. So does starting a fresh chat with a short summary.

Token counts here are rough estimates (about 1.35 tokens per English word). The conversation here is deliberately tiny so you can see the effect. Real windows hold tens of thousands of tokens or more, but the same thing happens when a long chat outgrows them.