The model has no memory unless you give it one

Outside of the active conversation buffer, the model remembers nothing. Vega can't pull up "what we were discussing yesterday". there is no yesterday for her, only the tokens currently in context.

When you come back to a conversation, you are the memory. Re-establishing context takes 60 seconds of effort up front and saves five turns of confusion. The shape is always the same:

1. Who I am. One sentence, identifying yourself.
2. What I'm trying to do. The goal, in one sentence.
3. What we already decided. The state the prior conversation left things in.
4. Where I'm stuck. The specific blocker.
5. What I need from you. The exact next question.

Skip any of these and the model has to guess. Guessing is where hallucination starts.

This is also the right shape for handing off to a human teammate, a support agent, or a coworker after a long weekend. You're not formatting for the model. you're formatting for anyone who lost the thread.

Your exercise

On the right: five labeled slots, five facts in the bank. Drag each fact into the slot it belongs in. When you click Check, Vega verifies the recap is complete enough to act on.