You will be able to explain what the context window holds and why the model remembers nothing outside it.
On Monday you spent twenty minutes with an assistant drafting a leave policy for your team. You told it the company has forty staff, most of them on shifts at a warehouse in Tuas, and that the boss hates long documents, and the draft it gave you was good. On Wednesday you open a new chat and ask for a matching overtime policy, and the reply is generic. It has no idea about the forty staff, the shifts or the boss.
Nothing broke. The assistant never knew those things in the way you know them. It only ever saw them while they sat in front of it, and that space has a name.
The context window is all the text the model can see at one time when it writes a reply. Lesson 3.2 showed that the model writes by reading everything so far and predicting the next token. "Everything so far" is the context window. If something is in it, the model can use it. If it is not, the model has no access to it at all.
Most people assume the window holds only what they typed. It usually holds a good deal more:
Hidden instructions from the company that built the app, which you never see, setting out how the model should behave, what it should refuse and what tools it has. Your own custom instructions or saved preferences, if you set any. Notes the assistant saved about you from earlier chats, if it has a memory feature. The whole conversation so far, with every reply the model wrote as well as your messages. Files you uploaded, or the text pulled out of them. Results from a web search or other tool, pasted in before the model answers.
Each new reply is written with all of that in view. That is why an assistant can refer back to something you said ten messages ago, and why it can follow the house rules its makers wrote without you ever asking.
This is where the Monday-to-Wednesday problem comes from. Lesson 2.1 described a trained model as a file of numbers, its dials, fixed when training ends. Chatting with it does not change those numbers. When you correct it, it does not get better at the task in general. It simply has your correction in its window, so the next replies in that same chat take it into account.
Close the chat and the correction goes with it. Start a new one and the model is the same file of numbers it was before you ever met.
There is one thing that can blur this. Depending on your settings and the provider, your conversations may be used to train a future version of the model. If that happens at all, it is months later, in a separate training run where your chats are mixed with huge amounts of other text. Don't count on it to remember anything you said.
Some assistants now remember things across chats. You mention that you are vegetarian or that you work in procurement, and weeks later it uses that fact without being told, which looks a lot like learning but works quite differently.
What happens is plainer than it seems. The app saves short notes about you, either because you asked it to or because it judged something worth keeping. When you start a new chat, it places some of those notes into the context window, alongside the hidden instructions, before you type a word. The model reads them the same way it reads anything else in the window. The dials have not moved.
Knowing this helps in two ways. First, you can usually open the memory settings and read, edit or delete what has been saved, because it is just text. Second, a wrong note will quietly steer every later chat. If it saved that you work in Johor when you moved back to Singapore last year, it will keep using the old fact until you fix the note.
Context windows are measured in tokens, the chunks of text you met in lesson 3.1. A long report, a contract, a transcript of a two-hour meeting: each uses up a share of the window, and so does every reply the model writes.
The size differs between products, and often between the free and paid plans of the same product. It also changes as companies release new versions, so any figure you read in an article may already be out of date. If you need to know, check the current number in the product's help pages, and remember that the hidden instructions and memory notes take some of that space before you start.
A bigger window lets you work with longer material in one go, but it does not mean the model reads every part of it equally well. Lesson 5.2, Why long chats drift and big files get skimmed, picks that up.
For now, the useful habit is to stop thinking of the assistant as something that knows you, and start thinking about what is in front of it right now. Open the assistant you use most and look at it with that question in mind. Some of what it can see, you typed. A fair amount, you did not.
List everything that might be inside the context window when you ask a question in your usual assistant, including things you did not type yourself.
Junxiong-WFG Organisation is an authorised representative of AIA Financial Advisers Private Limited (Reg. No. 201715016G).