The belief
You tell your AI assistant (a chat tool such as ChatGPT, Claude or Copilot) that your financial year starts in April. It replies "Got it." The next day, in a new chat, it plans everything from January.
Most people decide the tool is forgetful, or broken. The belief underneath is that it learned something from you and then lost it. That is a natural thing to believe. Everyone you have talked to learns from what you say, and a chat window feels like talking to a person.
It does not work that way. Behind the chat window is the model: the program that reads your text and writes the reply. How the model works decides what you can and cannot expect from it.
Three things people call learning
People use the word "learning" for three different things. Only one of them changes the model, and your chats are not part of it.
| What it is | When it changes | How long it lasts | |
|---|---|---|---|
| Training | The reading the model did before it was released: books, websites, code | Only when the company releases a new version, months apart | Until the next version |
| Context | Everything in the chat you are in: your messages, its replies, any file you attached | Every time you send a message | Until you close the chat. A new chat starts empty |
| Memory | Short notes the app saves about you and adds to the start of your chats | When the app saves a note, or you edit one | Until you delete it |
Training is finished before you type a word. Nothing you say changes it. The model you use today is the same one everyone else uses today, and it stays that way until the company replaces it with a new version.
Read the picture from the top. The notes, the chat and your new message go in. A reply comes out, and the model is exactly as it was before. That is how one AI assistant can serve millions of people at once without any of them changing it for the others.
What looks like learning
Inside a chat, it seems to learn. Tell it to keep answers short and the next answer is short. It has not learned anything. Each time you send a message, it reads the whole chat again from the top, including your instruction. The same thing explains the opposite problem: in a very long chat, the earliest messages can drop out of what it reads, so an instruction from an hour ago stops applying.
Across chats, memory carries notes forward. ChatGPT and Claude both have a memory setting. They save short facts about you, such as your job or how you like things written, and add them to the start of new chats. Those notes are ordinary text. You can read, edit and delete them in the app's settings, which tells you what they are: a list the app shows to the model each time, not something the model took in.
Your chats can still be used for training, later. Some plans (the free or paid account you signed up for) let the company use what you type to help build a future version of the model. Whether yours does depends on the plan, and we cover it in Is your data private when you chat with AI?. Even then it is a future version. The one you are talking to does not change.
What to do instead
You cannot teach it. You can brief it, which means giving it the facts it needs, each time. The question is where that briefing lives.
| If you want it to... | Do this |
|---|---|
| Use a fact in this conversation | Type it in the chat. It re-reads everything above |
| Know the same thing in every chat | Put it in the app's memory, or in its custom instructions (a settings box where you write what it should always know), then check what was saved |
| Work from your documents | Attach the files, or add them to a project (a saved workspace in the app that keeps your files and instructions together). It reads them each time. It does not take them in |
| Become a different model | That is called fine-tuning: retraining a copy of the model on your own data. Developers do it, at a cost. It does not happen by chatting |
Back at your desk
- Find the memory setting and read itIn the app's settings, look for Memory or Personalization. If it lists notes, read them. That list is everything the app carries between chats.
- Open a new chat and ask what it knows about youThe answer comes from those notes and nothing else. Anything you told it in an old chat that is not in the list is gone.
- Write your opening paragraph onceYour job, your house style, anything it keeps getting wrong, such as the April start to the year. Save it in custom instructions, or paste it at the top of each new chat.
- Correct it in the chat and move onFix the answer where it happened. Do not expect the correction to follow you into the next chat.
You are not teaching it. You are briefing it, every time.
When you know a thing, to hold that you know it; and when you do not know a thing, to allow that you do not know it; this is knowledge.