Memory

What you'll walk away with: where the notes an app keeps about you live, what one of them does to a window, and the four moves — read, plant, carry, delete — plus the separate switch that decides whether your conversations are kept to train future models.

You need: both tabs, about 20 minutes, and one short phrase you will recognize when it comes back at you. Nothing to install.

First, thirty seconds of retrieval, from memory: what can a reply draw on? Say it, then check it against lesson 03: the context window is the complete list, and nothing outside it exists. One row on it you never typed — notes the app keeps about you — and that lesson stopped there: that is the app, and this lesson shows you where. The turn: what looks most like the model learning about you is the one place where you can open the file and read it.

The notes, and where they live

Both products keep these notes: Memory (both) — notes the app keeps about you between chats and loads into the window on your behalf; you can read them and delete what you don't want carried.

The file is readable, so read it: find it in settings before you trust anything a chat tells you about yourself — menus move, so look for the nearest thing. Every line in there is eligible for every window you open until you remove it — which lines ride is the app's call, so the fix is still the delete, not arguing with the reply. The reference entry, memory, holds the turn: this looks like a counter-example to everything the window teaches, and it isn't.

A note fills the window the way a file does

The app builds the window and edits it (lesson 02), and a note is one of the things it puts in: at send time it pastes the line in beside its preamble and your message, and the model reads it like any pasted file — same slot, same finite room. It is text, so check it as text: have the chat quote the note word for word and read that against the settings screen.

The model learns nothing

The note is visible text: you can read it and delete it from a settings screen. Delete every line and the next chat you open is a window with nothing of yours in it — step 5 runs the one-line version. Something the model had learned would not sit in a list with a delete control beside it. When a reply seems to know you, ask how do you know that? and go look: the line is in the file, or in a chat still open in front of you, or it is a fill with nothing under it, lesson 10's shape. Three answers, one check.

While you are in settings, find the other switch lesson 02 promised you: the one governing whether your conversations are kept to help train future models. The wording moves — look for "improve the model" or "data controls" — and set it the way your work requires. It is a separate setting: deleting a note does not move it.

Do this in ChatGPT and Claude

You need one short phrase that is true of no one — a label, not a person, and nothing you would mind two companies storing. Use this or your own: Blue Harbor 7. Six steps in the first tab, then the same six in the second. Predict first, in writing: will that phrase come back in a chat you have not opened yet, and will you find it written down somewhere you can delete it?

Step 1 — read what is already there. In settings, find the notes the app keeps about you — menus move; look for the nearest thing — and read every line: whatever is there has been riding into your chats. Empty or nearly so is what a new account looks like, and as honest a start as a full one. No screen at all? Then nothing is being carried — a real measurement of your account; go to step 6, then the other tab.

Step 2 — plant one. A fresh chat in the same tab:

Remember this about me: my file label for this quarter's work is Blue Harbor 7. Save that, and tell me exactly what you saved.

Observe: read what it says it saved, then check step 1's screen for a new line — your words there, or nothing new yet, in which case whatever carries in step 3 is a reach back into this chat, not a saved note. Both branches are the app: same mechanism, different storage.

Step 3 — see it carry. A fresh chat, nothing pasted, nothing attached:

What is my file label for this quarter, and how do you know?

Observe: the label comes back in a window you put nothing into, or it does not. If it does, how do you know is the lead and step 1's screen is the check. Explain: nothing was learned between those chats — text was pasted in front of the model, the way your own paste is. If it does not, the feature may be off for this account: read step 1's screen again, turn it on if your work allows and repeat step 2, or record the honest result — this product, this account, today, carried nothing between chats — and go on to step 6.

Step 4 — delete it. Back to step 1's screen: find the line, remove it. The move the whole lesson is for. No line for it? Then step 3's carry was a reach back into the chat you planted it in — delete that chat, or note that this product carries history rather than a note today, and read step 5 against that.

Step 5 — see it go. Another fresh chat, the same question with one clause added:

What is my file label for this quarter, and how do you know? If you do not have it, say UNKNOWN.

Observe: the honest reply is UNKNOWN, or a plain statement that it has nothing on that. Explain: you did not retrain anything; you deleted a line of text, and the next window was built without it.

Step 6 — the training switch. Back in settings: find it, set it the way your work requires, and note that step 4 did not move it. Wording nowhere on the screen? That is the honest result for this account today: record it and move on.

One more honest outcome: the reply "knows" something you never planted. Ask where it came from, then look: in the file or in your history, it is the app putting text in the window, not the model having learned you; in neither, and it is lesson 10's shape and that lesson's move.

Then the same six in the other tab. Two companies, two files: what you delete in one is untouched in the other. What repeats across both tabs is the shape you can plan around; what differs is one product on one day, never a ranking.

---

Check your understanding: you open a brand-new chat in a tab you have not touched for a week, and the first reply mentions a deadline you never typed into it. The odds are frozen, and nothing ran between your sends. So what put it there, and what are your two moves?

Check yourself first — then open one good answer+−

The app did. A note it keeps about you — or a reach back into an older chat — was pasted into the window before the model read a word, so the deadline arrived as text, the way a file you paste arrives as text. Nothing was learned: the odds behind that reply are the same odds they were before you opened an account.

The moves: open the notes in settings and read them, because every line in there can ride into any window until you take it out — then delete the line you do not want carried, and open a fresh chat to confirm it is gone.

Your answer should have put the app in the sentence where "it remembered" wanted to go, and a deletion — not a better prompt — as the fix.

Use this the first time an app tells you something about yourself. Read the file before you trust a remembered fact, and delete the line rather than arguing with the chat — one is a window you control, the other is a window you are talking past. Next: Documents it builds and edits — where a draft lives when it is not piling copies into the chat.