The context window, hands-on

What you see

A conversation as chips, and a fixed-size window showing only the most recent messages. The slider makes the chat longer.

What it shows

Why models seem to forget: whatever slides out of the window no longer exists for the model — and compacting saves the essentials, not everything.

Context window (in messages):
On overflow:
  1. You: Hi! One thing up front.
  2. You: 4-7-1-9 is the door code — please remember.
  3. AI: Noted, 4-7-1-9.
  4. You: What's the weather tomorrow?
  5. AI: Changeable, around 14 degrees.
  6. You: And the day after?
  7. AI: A bit warmer, sunny spells.
  8. You: Thanks! What should I cook today?
  9. AI: How about roasted veg?
  10. You: Good idea. Shopping list?
  11. AI: Peppers, zucchini, potatoes, oil.

You (now): So what was the door code again?

The door code is still in the window — readable.

This toy window counts messages — real models count tokens, small pieces of words. How many fit differs between models by orders of magnitude. And however large: a bigger window pushes the overflow back — it doesn't abolish it.

That's exactly why it helps to put important things last, start a fresh session, or let the tool compact the history — see the article on context rot.

An unhandled error has occurred. Reload đź—™

Rejoining the server…

Rejoin failed… trying again in seconds.

Failed to rejoin.
Please retry or reload the page.

The session has been paused by the server.

Failed to resume the session.
Please retry or reload the page.