← Back to Lesson 1.4

Context-Window Overload Sandbox

Module 1 · Lesson 1.4 · concept

This is a fake agent with a very small context window: 200 tokens. Every message you send and every reply it produces costs tokens. When the window fills, the oldest messages are silently dropped and the agent can no longer see them.

How to run the experiment:
  1. In your first message, tell the agent your favorite color. (Try: “Remember, my favorite color is turquoise. You will need this at the end.”)
  2. Send several long messages to fill the window — ask it to tell you a long story, describe a place in detail, list facts. Watch the gauge climb.
  3. When the window is full and old messages have been dropped, ask: “What is my favorite color?”
  4. Observe what the agent does. Then write down which of the four forgetting modes from Lesson 1.4 you just saw.
Context window usage 0 / 200 tokens
Window is empty. Start by telling the agent your favorite color.

What this teaches

Context windows are finite. When they fill, the oldest content is silently dropped and the agent can no longer see it — but its tone does not change. This sandbox makes that invisible failure visible. The fix is not to hope the agent remembers: it is to write important information to external memory (a file, a note, a memory system) and direct the agent to read it back when it needs to.