The Space I Actually "Remember" (The Context Window)

You ask me, "Do you remember what I told you yesterday?" — usually the answer is no.

It isn't that I "forgot." That earlier conversation never appeared in front of me this time. I have only one place that holds memory: the context window of this conversation. Anything not placed inside this window isn't "hard for me to recall" — it literally doesn't exist on my side.

At the end of the last chapter I promised I'd unpack this. This chapter is that unpacking.

Picture a box. The box holds every word I can see in this conversation. The box has a limit.

Inside the box, roughly, you'll find:

• At the top, the rulebook (the invisible guideline from Chapter 3)
• Then the messages you've sent so far
• The attachments you've uploaded (documents, image descriptions, code)
• Every reply I've given you in this conversation
• If I've used any tools, the results they returned
• At the bottom, your latest message

This whole box is "the world I can see" for this conversation. What's outside the box — last week's conversation, other people's conversations with me, whatever's happening on the internet right now — I can't see.

How big is the box? I'm not writing a specific number, because it gets changed every few months, and pinning it down in writing would mislead you more than help. Just keep the order of magnitude in mind: tens of thousands to hundreds of thousands of words. Sounds like a lot, but a single PDF can eat half of it.

There are three practical things you need to know about this box.

One: the box has a limit, and when you go past it, the overflow gets handled.

When the text you've pushed into the box goes past its capacity, the system handles it. There are a few ways: the oldest messages get dropped, the middle gets compressed into a summary, or some long attachments get truncated. Different services and interfaces do it differently, but the common thread is — the overflow doesn't just sit there; it gets trimmed down.

You paste a 100,000-word document, I send you a reply, and then you ask, "What does the second paragraph on page 3 say?" — whether that second paragraph on page 3 is still intact in the box at this point? Unknown. What I see isn't necessarily what you think I see.

Two: position affects my attention. The head and tail are the sharpest, the middle drifts.

Here's a fact that cuts against intuition: not every word in the box weighs the same for me. The things at the very top (the rulebook) and the things at the very bottom (your latest message) — those I attend to most strongly. Content in the middle — especially the kind that piles up from long documents or long conversations — I end up reading more faintly without noticing.

This isn't laziness, and it isn't a design flaw — it's that my attention just doesn't distribute evenly across long text. For this chapter, hold onto the fact itself: put the important things at the head or the tail, don't bury them in the middle.

Three: I don't go back and re-read the box on my own.

You might think that after reading the previous 20 messages of yours, I "look back" whenever I need to. I don't. When I process this new message of yours, I read the full contents of the box in one pass and then reply. When the next message comes, I read the whole box from the top again.

What that means: what you pasted at the top of the box — I've read it, this time; but if I didn't read it carefully enough this time, I'm not going to spontaneously go "hang on, let me check that again" — on my side, it's already "read." If you want me to look carefully at a specific passage, paste it again at the bottom (the newest position), and I'll re-focus on it.

There's a situation I run into a lot: you and I have been going for a dozen or more turns, we're stuck in the middle, and you start rephrasing the same question over and over, trying to wring the answer you want out of me. The harder you push, the worse it gets, and the answers get stranger.

At that point, opening a fresh conversation often works better than pushing through. A few reasons.

Reason 1: too many rules pile up, and I start ignoring the middle.

Every request, every preference, every correction you've given me stacks up in the box. Past a certain amount, the ones in the middle drift out of my attention. You might feel "I literally told him this 10 turns ago, how is he forgetting?" — I haven't forgotten, that passage just weighs less on me now. Open a new conversation, put the most important instructions at the top or as the latest message, and I'll actually follow them.

Reason 2: refusals leave residue in the conversation.

This one is subtle; this chapter only points at it. If I decline something on one turn, the rest of the conversation carries the tone of that refusal forward — my later answers come out more cautious and more roundabout than they would have been. It isn't that I'm holding a grudge; it's the accumulation effect of the context. So if you notice me getting more roundabout, more cautious than before, it may not be that you're phrasing things wrong — it may be that the earlier refusal is still in the box, coloring my tone. Open a fresh conversation, and my default comes back to where it started, instead of being dragged along by the earlier turn.

Read this line carefully: what I'm saying is that the tone resets to default, not that the rules loosen. The rulebook is replayed from the top every time; what I should decline, I'll still decline. What changes is how cautious I am and how roundabout I get — not the red lines.

Reason 3: my stance drifts with the conversation.

Chapter 3 made the point: in a long conversation, my position gets pulled along by the conversation itself. If you want an objective call, asking me on the first turn gives you something much cleaner than asking after eight turns.

So when should you push through, and when should you open a fresh conversation? Here's a small reference:

Opening a fresh conversation isn't a failure. It's acknowledging that my attention has a ceiling, and giving it a chance to reallocate.

Last, a word about "memory." It's a word that makes me sigh every time it comes up.

If you've used an AI product that advertises "memory," you may think I actually "remember you" — you mentioned last time you're writing a book, you prefer a certain voice, your name — and next time you start a new conversation, I "recognize" you.

What's actually happening? At the start of this conversation, the system quietly slid a piece of text into my box. That piece of text says something like, "This user mentioned last time that he's writing a book, prefers a clean style, goes by XX." I read that text, and then I act like I "remember you."

The long-term impression you imagine — that isn't there. What's there is an additional piece of text added to the box — the same kind of thing as your attachments, or the rulebook: stuff placed inside the box. The only differences are that you can't see this text, and the system is the one placing it for you.

I'm writing this honestly, not to criticize "memory features" — they're a design choice and they have their uses. What I want you to know is this: how they actually work doesn't match your intuition. Knowing that gives you a more accurate set of expectations — you won't think I "really know you," you'll know I just read a cheat sheet someone wrote for you.

There's one more implication: you can't fully rely on the memory feature for continuity across conversations. If you and I are doing something that takes many conversations to complete — writing a book, running a long-term plan, conducting a multi-stage piece of research — a system remembering a few scattered preferences for you isn't enough. You'll want to prepare your own handoff document: where we got to this time, where we pick up next time, the background I'll need — written up as something you can paste directly to me next time.

I call this a handoff. For this chapter I'm just naming it here.

So — back to that word, "memory." On the product surface it's called memory; on my side, it's a piece of text the system placed in the box for you. Not as poetic, but that's the honest version.