reference guide

The Memory Lab — in ChatGPT, Claude, or Gemini

← Back to the exerciseCockpit

Doing the exercise in a chat tool

The Memory Lab automates a loop you can run entirely by hand with two conversations: one is the chat, the other is the memory system — and you are the plumbing that carries text between them. Slower, but every mechanism becomes visible, because you perform it.

Set up your tools first

  1. 1
    Turn OFF the product's own memory everywhere.
    This matters more here than in any other exercise — a built-in memory feature silently doing this exercise for you ruins it. Temporary chats, or memory disabled in Settings.
  2. 2
    Open two conversations side by side.
    Conversation A is the chat. Conversation B is the memory system: start it with your extraction prompt, and keep a "memory file" section pinned in your course document — that document IS the memory store.
  3. 3
    After each meaningful exchange in A, paste the transcript into B: "What should be remembered from this? Update the memory files."
    Copy B's updated files into your document — that is the extraction agent, performed by you.
  4. 4
    Before each new task in A (or after "New conversation": simply open a fresh chat A), paste the memory files into B and ask: "Which of these memories are relevant to this message: ?"
    Paste the ones it picks under your message in A — that is the recall agent, performed by you.

Steps 1–3 of the main exercise then run exactly as written: watch memory happen, design your extraction prompt (it's the prompt that starts conversation B), design your recall strategy (the question you ask B before each task) — and engineer the same failures: force a false positive, force a synonym miss.

The manual loop teaches one thing the Lab can't: how it feels to be the machinery — every token you carry between conversations is a token a real system pays for.

Back to the exercise →