The Third Umpire · Lesson 2 — The Notebook ← Course

The notebook — memory across chats.

Open a new chat and the screen is wiped clean. So how does an assistant ever “remember” you at all?

The last two pages showed a whole conversation sharing one window. Here's the part that surprises people: the moment you close a chat and open a new one, that window is wiped completely clean. The new chat has no idea what you said yesterday — or a minute ago, in the chat you just closed.

So by default, an assistant does not remember you between chats. The review monitor is wiped at the end of every match, and whatever was on it is gone. To recall a batter's habit next season, the umpire jots it in a notebook and carries it between matches.

That notebook is the trick. Some assistants keep a small store of chosen facts outside the chat window — a name, a preference, a detail or two — and it survives the wipe. At the start of a new chat, a few are quietly loaded back into the fresh window, present from the very first message.

The key idea is that these are two different things. The window is temporary working memory, wiped every chat. The notebook is deliberate, lasting memory — a short list of saved facts. Only the second one carries over.

Three facts are already in this chat. Choose memory off or on, start a new chat, and see what survives the wipe.

Step 1 of 4
Two umpires, same match — one with memory off, one with memory on.
Memory OFF
notebook — off
Memory ON
notebook — empty
Your first chat. You have told the assistant three things about yourself.
Memory · kept between chats
saved on the way out · loaded back on the way in
This chat — your first conversation

Memory is just a list, loaded back in.

Remembering is not the model holding on to your old chats. It is a short list of facts, kept on the side and re-loaded into each new window at the start. 

# a new chat starts with an empty window — it remembers nothing on its own
system = briefing

# memory is a SEPARATE store that lives outside any single chat
memory = load_saved_facts(user)        # e.g. ["Arjun", "left-arm spin", "Hurricanes fan"]

# at the start of a new chat, saved facts merge into the ONE system message
if memory:
    system += "\n\nAbout this user: " + "; ".join(memory)

window = [ {"role": "system", "content": system} ]

# only now does the new chat "know" you — the facts were re-loaded, not remembered
reply = umpire.chat(window + [{"role": "user", "content": message}])

Two kinds of memory.

The chat window is working memory — wiped clean the moment a conversation ends. An assistant only recognises you across chats if it keeps a separate notebook of saved facts and loads them back into each new window. No notebook, no recognition.

So “remembering” is not the model clinging to your past chats. It is a small, deliberate list of facts, saved outside the window and re-loaded at the start — something you can usually see, edit, and delete.

Next: the maths →
▸ Go deeper — how memory really works optional

The model itself does not change when it remembers you

Saving a fact about you does not retrain or rewire the model. It just adds a line of text to a store. The model that answers is the very same one everyone else is using; your notebook is the only part that is yours. (Think back to Lesson 1: the umpire's training — everything it absorbed across seasons of past cricket — is baked into the model for everybody. Memory is a thin, personal layer laid on top.)

You can usually see it, edit it, and switch it off

Because memory is only a list of text, there is nothing mysterious to undo. Good products let you read exactly what has been saved, delete any of it, and turn the whole feature off. What you can read, you can remove — which is worth knowing in a time when “the AI remembers me” can feel unsettling.

The same trick, pointed further

Loading a few saved facts into a fresh window is a small version of a move you'll meet again: the same trick, pointed at documents instead of facts about you, gets its own lesson later.