1. 1. Words are not the unit
  2. 2. Split a sentence yourself
  3. 3. The language tax
  4. 4. Check yourself
  5. 5. What a context window is
  6. 6. Fill the window
  7. 7. Long isn't the same as read
  8. 8. Check yourself
  9. 9. Weights, not a diary
  10. 10. Memory features, honestly
  11. 11. Final check

Lesson 10 of 11 · Part 3 · Read · 4 min Draft

Memory features, honestly

What an app's memory feature really does: it saves notes about you and puts them back into the window.

The filing cabinet next to the model

Some chat apps can bring up things you told them weeks ago. The last lesson showed the model itself cannot do this. Its weights stay fixed. So something else is doing the remembering.

Picture a front desk staffed by a new temp every morning. The temp has never met you. But there is a card file on the desk, and before you walk in, someone puts your card on top. The temp reads it and greets you by name. All the remembering lives on the card.

Hello, AI describes memory features the same way. Many chat products add a separate memory feature that saves notes about you and quietly pastes them into new chats. The book calls it a filing cabinet next to the model. It is useful. It is also separate from the model learning anything.

It all comes back to the window

Once you see the cabinet, the rest follows from Part 2. Saved notes help only when they are put into the context window. Once there, they are text like any other, and they take up room measured in tokens.

In effect, the app does what you could do by hand: paste a reminder about yourself at the top of each new chat.

A note the app leaves out is invisible to the model, just like a message that has scrolled away.

Without such a feature, Hello, AI is clear: a new chat starts from zero. Yesterday’s conversation is gone unless you paste it back in.

When it feels like a friend

A chatbot that remembers last week’s plans can feel personal. Hello, AI’s chapter on companionship takes that feeling seriously.

The warmth comes from machinery you already know. It is next-word prediction, tuned on millions of human ratings of what a kind answer looks like, aimed at your context window. The book’s phrase for it is “genuinely something, and genuinely not someone.”

Memory can also be switched off. Hello, AI reports that OpenAI’s parental controls for ChatGPT let a parent turn off memory, along with voice mode and image generation, on a linked teen account.

In one sentence: an app’s memory is saved notes placed back into the window, and the model itself still learns nothing from your chat.

Quick check

One question to lock it in. It doesn’t count against you.

How does an app's memory feature remember you between chats?

Sources