Free sample lesson ยท Why did it forget what I said earlier?

Working Memory, Not Storage

It lost the instruction you gave at the top of a long conversation.

Lesson 1 of the Wondering course Why did it forget what I said earlier?

The instruction that vanished

Early in a long conversation you said: keep everything under 200 words and never use bullet points.

It worked for a while. Forty exchanges later you are getting 600-word answers full of bullets.

It did not decide to ignore you. From where it is standing now, you never said it.

Tokens, not words

Models do not process words. They process tokens: chunks of text, usually smaller than a word.

A common word like the is normally one token. A long or unusual word splits into several. Punctuation and spaces count.

A rough conversion for English is that a token is about three quarters of a word, so 1,000 tokens is roughly 750 words. That is only a guide, and it breaks down for code, other languages, and numbers.

This is why every limit you see is stated in tokens rather than pages, and why a page of dense technical text costs more than a page of ordinary prose.

The window

There is a maximum number of tokens the model can consider at once. That is the context window, and it holds everything: your instructions, the whole conversation, any documents, and the reply being written.

The crucial word is working. This is not storage. It is what is in front of it right now.

Anything outside the window is not forgotten in the sense of having been known and lost. It was never visible at all.

Every turn is rebuilt from scratch

This is the part that surprises people.

There is no running conversation held somewhere. Each time you send a message, the entire conversation is re-sent to the model, and it reads the whole thing fresh.

That is why it can refer back to something from twenty messages ago: that text is still being included. And it is why the beginning eventually disappears: at some point the conversation no longer fits, and the oldest material gets trimmed to make room.

When your early instruction stopped being followed, it had fallen off the edge.

What this predicts

You can now anticipate several behaviours:

  • Long conversations drift, because the early framing is gone
  • Uploaded documents may be too large to be fully present
  • New chats start blank, which is a feature when the old thread is cluttered
  • Restating matters, and is not redundant

The practical habit: for anything long, repeat the critical constraint near your latest message. Not because it did not understand the first time, but because the first time may no longer exist.

Keep going

This is lesson 1 of 2 in Why did it forget what I said earlier?. Next up:

  1. The Middle Gets Lost

Start this course free

Wondering turns any topic into short lessons you can read, watch, or listen to, with practice questions and an AI tutor.