The Invisible 98.8%

I spoke to an AI.

"How about building an agent that captures a MacBook screen and talks about it?"

A short sentence. But when it reached the AI, it did not travel alone. I opened the actual input the AI received. My words were about 350 tokens. The full input was about 28,000 tokens. A ratio of 1 to 79. My words were 1.2% of the total.

What was the other 98.8%? A 6,000-token system prompt. An 8,000-token tool specification. 2,500 tokens of auto-injected skill instructions. Plugin lists, dates, hook results. Things I did not write, did not see, did not know about. I had said one sentence, and the AI was receiving a short novel's worth of instructions alongside it.

The asymmetry was uncomfortable. I thought I was talking to an AI, but in reality I was sending a small signal on top of a massive system. Humans cannot look inside their own heads. But the inside of an AI's head is data. It can be opened.

So I started building a visualization tool. First I tried color coding. System in red, tools in yellow, user in white. Pretty, but flat. Color is packaging that hides information. I dropped color and kept only typography. Size and brightness. My words are large and bright. The system prompt is small and dim. No labels, no icons. Seen that way, for the first time, I could see.

Move the slider below. This is the actual context the AI received while I was writing this essay.

The top of the screen is a wall of tiny, barely visible text. English, dense and mechanical. Past that wall, suddenly, large clear Korean appears. My words. A small island floating in an ocean of machine text. What I felt was less discomfort than awe: this tiny signal was steering the direction of that enormous machine.

Here the question changed. People usually say to save tokens. I came to think the opposite.

Suppose the AI's maximum context is 200,000 tokens. What I am using now is 28,000. That is 14%. The remaining 86% is empty. An empty context is an empty mind. The analogy is imperfect, since the AI's knowledge already lives in its model weights. But insofar as the context determines behavior at this very moment, emptiness is close to drifting without direction. The question is not how to save but how to fill better. The problem of assembling the best team on a fixed budget.

I ran an experiment. I inserted the same instruction three times. The result changed. What the model had ignored with one copy, it now followed precisely. An AI distributes attention across the parts of its input. If the same content sits in three places, attention goes to three places. Three times the presence. Without changing the model, changing only the input changes the behavior.

Here I make a leap. I am not certain. Perhaps long-term memory is a function that decides how often to place a given piece of information in the context. Important things appear frequently; less important things get pushed out. It will not be all of memory. But frequency, as one axis, works. A 2025 study from Google Research shows this.¹ Prompt repetition alone improved performance on 47 of 70 tests, with zero degradation.

Filling. This is the first half of memory.

Filling has a limit. What happens when 200k is full? You have to choose. What to keep and what to throw away. And here I met a problem I had not anticipated.

"Go ahead."

I said this to the AI. Over a session that ran 29 days, I said it many times, and each time it meant something different. At one point it meant: now implement. The design was done; only the code remained. At another point it meant the opposite. Just before, I had said "this is bad," "roll it back." That "go ahead" was not about implementation but about review.

The current input is the same. The correct action is different.

When a conversation grows long, you cannot hold all of it. You have to summarize and compress. But what do you throw away? Most memory systems work by "find past moments similar to the current question." Retrieval. But retrieval cannot tell the two cases of "go ahead" apart, because the query is the same. What is needed is not retrieval but preserving the boundary — the line where the same input demands a different action.

Two pasts exist, the current input is the same, and the correct action differs. The "go ahead" that follows a past of relentlessly doubting the approach calls for reconsideration; the "go ahead" that follows a past of approving the design down to the unit calls for implementation. If a memory system collapses these two pasts into one, the action breaks. Call this history aliasing. There is also the failure in the opposite direction. "Be more rigorous," "cut the nonsense," and "isn't this wrong?" look different on the surface, but the action they demand is the same. Splitting these three into separate states makes memory sensitive but not useful. Good memory does both at once. It separates pasts that must lead to different actions, and it merges pasts that may lead to the same one.

Seen this way, memory is a compression function. You cannot live carrying every past, so you have to reduce. Whether the form is a retrieval set, a summary, or a state variable does not matter. What matters is which pasts the representation merges and which it separates. Good compression is not about making things small. It is about keeping the differences that change action and discarding the rest. It sounds obvious, yet most current memory systems are not evaluated by this criterion. They measure how much is remembered, not which boundaries are preserved.²

Here filling and folding meet. What to put in when filling the context, what to keep when it is full. Both are the same question: does this change a future action?

One thing keeps nagging at me. To find similar contexts, most systems convert text into vectors and measure distance. That is a point. A snapshot of a moment. But the essence of context is sequence. "It's okay" after comfort is not the same as "it's okay" after a fight. "Go ahead" was the same way. Its meaning depended on what had flowed in before it. If the point is the same but the trajectory is different, the thing is different. To preserve boundaries, you have to compare trajectories, not points. This is still an open problem.³

The session that became the material for this essay ran 29 days. What began as an attempt to build a screen capture agent turned into a visualization tool, passed through filling and through folding, and arrived at the question of memory. I ended up somewhere that was in no plan.

One thing I realized: the act of looking into the context itself consumes context. When you say "show me your context," that sentence is added to the context. Observation changes the thing observed.

When you talk to an AI, what you see is 1.2%. The other 98.8% is invisible. The invisible part decides more than the visible part. When it fills up, you have to decide what to keep and what to fold. That decision is memory.

Is this only about AI? When we talk to each other too, the other person sees only a fraction of my words. Why I said them, what experience made them — the other person cannot know. We are all speaking on top of an invisible 98.8%. The difference is that the AI's 98.8% can be opened. Ours cannot. Not yet.