AI, plainly

06 Lesson 6 of 15

How much it holds at once

A conversation is not a growing record. It is a fixed-size space, and things fall out of the back of it.

  • About 2 min read
  • Interactive 3D diagram
The version most people start with

It can see everything we have said. The conversation is right there on screen, so surely it can all be seen. If you mentioned something from twenty messages ago, it would know what you meant.

The idea to leave with

Either way, the early detail is no longer fully there.

See it in the lesson

Up to a point, yes. Each time you send a message, the tool gives the model the recent conversation plus your new message, and the model continues the whole thing. That bundle is called the context window, and it has a size limit.

When a long conversation goes past that limit, something has to go, usually the oldest part. Tools deal with this in different ways. Some quietly drop old turns, some quietly summarise them, and some warn you. Either way, the early detail is no longer fully there.

The diagram shows the idea. The large faint shell is what the model learned in training, which does not change while you talk. The small bright box is this conversation. Add turns and watch it fill up, then watch the oldest turn drop out of the bottom.

    Three separate stores, drawn to show how much each one matters: a large fixed one, a small one that fills up, and a detached one that stays dark until it is used. The conversation box here holds six turns. Real ones hold far more, but they still have a limit.
    Read this diagram as text

    The line under the buttons always says how many turns are in the conversation, how many fit, and how many have fallen out for good. The numbered list below it is the conversation as the model has it right now. Pressing Add a turn past the limit removes the oldest entry. Pressing Start a new chat empties the list, and changes nothing about what the model learned in training. With memory switched on, the new chat starts with a saved note about you already in it.

    What this predicts

    This explains two things you may have noticed. In a long conversation, it starts forgetting instructions you gave near the start: the tone you asked for, the format, the limits you set. And in a very long one, it may contradict something it told you earlier without noticing.

    The fix is simple and it works. For anything long, repeat the important instructions in your latest message. Do not trust that the early ones are still in view. And when a long conversation starts to drift, start a fresh one and paste in only what matters. Starting a new chat is nothing to feel bad about.

    One analogy. It is closer to a desk than a filing cabinet. Only what fits on the desk is in play, and putting something new down pushes something else onto the floor.

    Where the analogy breaks: you can pick things up off the floor. Here, once a turn has fallen out of the window, it is simply gone. It may still be on your screen, which is what makes this confusing.

    The move to make

    In a long chat you already have, ask it to repeat something specific you said at the very start. Whether it can tells you where the edge of the window is for that tool.