The conversation has a memory limit

What the model can see

The model does not have a memory the way you do. It has a window. Everything it can consider when writing its next word (your instructions, your pasted material, its own earlier answers) has to fit in that window at the same time. Anything outside it does not exist for the model.

The window is measured in tokens. A token is a piece of a word: roughly three-quarters of an English word, and often less than that for Hindi, Swahili or Arabic, which are split into more pieces. The big paid models offer windows in the range of 128,000 to a million tokens: 128,000 tokens is roughly 90,000 words, a 300-page book. Free tiers are typically given a fraction of that, and the tool does not always tell you the number.

For a teacher, the practical version is: it can hold a chapter. It probably cannot hold a textbook. A syllabus, a mark scheme, a class set of short answers: fine. Five past papers at once, a whole term’s resources, a 200-page PDF: it will either refuse, or quietly keep only part of it.

The quiet failure

The refusal is the good outcome. The quiet failure is worse: the tool truncates. It keeps the beginning or the end, drops the middle, and answers as though it had read the lot. You asked for chapter 7 questions; chapter 7 was in the dropped part; you get confident questions on chapter 3.

Nothing warns you. The only defence is to give it less and check that it has what you think it has. A cheap check: “Before you start, tell me the heading of the last section you can see.” If it names the wrong section, the material did not fit.

Long conversations forget their beginning

The window also holds the conversation itself. Every message you send and every answer it gives takes up room. Forty exchanges into a planning session, the instruction you gave in message one (“always use Kenyan shillings, sentences under 15 words”) has slid toward the edge of the window, and eventually out of it.

Some tools handle this by silently summarising the early part, which is why the model starts “forgetting” your constraints and reverting to dollars and long sentences. It is not being careless. Your instruction is no longer in front of it.

Two habits fix this:

1. One task, one chat. Worksheet in one chat; parent letter in another; misconception list in a third. Long threads drift.
2. Repeat the constraints when they matter. Rather than trusting message one, paste the two lines of constraints again with the request that depends on them.

Give it the material every time

Because there is no memory between chats, the model does not remember the syllabus you pasted yesterday. Tomorrow’s chat starts blank. This feels wasteful, and teachers try to work around it by keeping one enormous chat open for weeks, which produces the drift above.

The better workaround is a text file on your phone with the things you paste often: your class description, the syllabus extract for this term, your standard constraints. Copy, paste, ask. Twenty seconds. Some tools offer a way to store this permanently (custom instructions, a “project”, a “Gem”) and a later lesson covers using those. But the plain text file works everywhere and never gets lost when a product changes.

Images and files cost more than you think

A photo of a textbook page costs several hundred to a couple of thousand tokens, depending on the tool. A scanned PDF of thirty pages can be thirty images. It adds up faster than text, and it hits the free-tier limits sooner. If you can type the two paragraphs you need, do; it is cheaper and more accurate than a photograph of them.

The limitation

Even inside the window, attention is not uniform. Models are measurably worse at using material buried in the middle of a long input than material at the start or end: researchers call it “lost in the middle”. So if the one constraint that matters most is in paragraph nine of your pasted brief, move it to the top or the bottom. Where you put things changes what gets used.