What memory forgets
I never learned calculus.
I memorized it. Limits, volumes of revolution, integration by parts. I didn't understand what a limit was. I learned what the problems looked like and what steps followed. Pattern in, answer out.
It took enormous effort. Cramming before exams, hoping the patterns would stick long enough to survive the test. They usually did. Sometimes they didn't. Chemistry was worse. Fewer patterns, more noise, and my system started falling apart.
I passed. But I never learned it.
Programming was different.
I started coding at fourteen, and nothing about it felt like school. I didn't memorize syntax and replay it. I understood what a loop did, and from that I could write any loop I needed. I didn't store answers. I built machinery that produced them.
I didn't know why one felt like dragging weight uphill and the other felt like breathing. I just assumed I was bad at math and good at computers.
It took me years to realize it had nothing to do with talent.
What I was doing in calculus was brute-force storage. Every problem was a separate entry in my head, filed by shape, retrieved by recognition. Expensive to store. Fragile to recall. And completely useless the moment a problem showed up in a shape I hadn't seen before.
What I was doing in programming was compression. I wasn't storing answers. I was storing rules that could regenerate answers. A loop isn't a thing you memorize. It's a concept you understand once and reconstruct forever.
The difference between those two modes is the difference between memorizing a language and speaking it.
This is how all expertise works.
An experienced doctor doesn't remember every patient. They've compressed thousands of cases into heuristics. Young guy, chest pain, probably not cardiac. That's lossy. Sometimes dangerously so. But it's fast, cheap, and right often enough to be useful.
An expert chess player doesn't memorize more board positions than a beginner. They encode them in better chunks. The information compresses for free because they have a structure to hang it on.
Experts don't remember more. They compress better.
Now give an AI a memory and watch it do the opposite.
ChatGPT remembers me. It keeps a list. "Johnny is a software engineer." "Johnny likes running." "Johnny ordered pizza once." Flat facts, no hierarchy, no structure. Every detail stored at the same priority. It's my calculus brain. Pattern in, fact out, no model underneath.
It feels wrong the same way it would feel wrong if a friend did it. You're dialing in your espresso and they hit you with "well, as a software engineer, you might prefer a lighter roast." You mentioned your job once. Now it follows you into every conversation. That's not memory. That's a fact the model doesn't know what to do with.
ChatGPT's memory is cramming for the test. It works until it doesn't.
Think about how you actually know someone.
You don't remember every conversation. You remember a model of the person that those conversations built. Recent things are vivid. What they told you yesterday, what they're stressed about this week. Older things are compressed. Where they grew up, what they care about, how they think. And somewhere deeper, there are a few core things you just know about them. The stuff that almost never changes.
You also forget. Constantly. Not by accident, by design. Your brain can't afford to store everything, so it stores what matters and lets the rest go. Every night when you sleep, your brain replays the day, decides what transfers to long-term storage, and prunes the rest.
You don't wake up remembering the transcript. You wake up with the updated model.
That's what AI memory should look like.
Not a list. A model. Layered, consolidated, lossy on purpose. Recent context in high resolution. Long-term understanding that fades slowly. Core insights that almost never change. And the willingness to throw away everything else.
The cost of remembering everything is that nothing has weight. The power of forgetting is that what remains actually matters.
I think about my calculus brain sometimes. All that effort, all that cramming, and what I actually retained is almost nothing. A vague sense of what a limit looks like. The word "integral." Shapes.
But I still write code every day. Not because I memorized more, but because I compressed better. The machinery is still there, still producing answers, years after I built it.
The best memory isn't remembering more.
It's knowing what to forget.