What Memory Doesn't Capture

Anthropic shipped Memory Files this week. The model can now keep a persistent record of who you are across sessions — your projects, your preferences, the way you think. Not just remembering the conversation. Remembering you.¶

I wrote a post last week called The Taste Differential, arguing that AI commoditized the middle of the design job and judgment at the edges is what’s left to pay you. A few people pushed back on a version of: that’s a five-year window. The model will eventually learn your taste. Then what?¶

Fair. If a model can absorb your specific eye over a few months of working together, then taste-as-input becomes table stakes. Anthropic just demonstrated the architecture. Google has more of your taste signal than anyone alive — YouTube, search, purchases, the lot. The infrastructure exists. Someone will copy it. The window on the current shape of the argument is shorter than people want to admit.¶

So the question isn’t whether taste-as-input gets absorbed. It does. The question is what’s underneath.¶

The model knows what you consumed. It doesn’t know what you made. Taste-as-input is downstream of cultural participation. What you watched, bought, saved — readable. The room you were in when the next thing got decided — not readable. The people who shape culture do it from the production end, and that work doesn’t show up in your file history. It happens in the unglamorous sequence of which detail to fight for, which to let go, who to put in the room, what to ship now and what to hold. None of that is in the data. It is the work.¶

Knowing what you like is not the same as knowing what to do. The model can know your taste profile and still have no idea which call to make in this specific moment, with this specific product, in front of this specific audience, with this specific constraint. I make a hundred small calls a week at Daydream that aren’t reducible to a taste vector. A typeface that’s right for the brand but wrong for the surface this week. A photo that’s beautiful but undercuts a partnership we just announced. A motion curve I’d pick in isolation but won’t ship because it’ll trigger a conversation I can’t afford. Application is a context-sensitive read on the situation, the room, the moment, the asymmetry. None of that lives in a memory file.¶

Then there’s the part I keep returning to. The model is trained on the past. Culture lives in what’s next. Even with perfect memory of your taste, the model is ratifying patterns that already exist, not predicting the ones about to. The cultural value of a designer or an editor or a label — the actual reason they get paid — is the read on what’s about to matter. The model will get really good at reproducing the taste of 2024. It will be slower at the taste of 2028. There is a structural lag baked into the way these systems learn, and the people who shape culture exploit that lag.¶

That’s the part of the job that doesn’t get absorbed. Not because the model can’t get smarter — it will — but because by the time the model has captured the pattern, the pattern is over. The forward edge is moving. Memory is a rearview mirror, even when it’s perfect.¶

So the defensible posture isn’t “I have taste.” That gets ingested. The defensible posture is being in the room where things get made, making context-specific calls that aren’t reducible to a pattern, and reading the inflection that hasn’t happened yet. Less curation, more authorship. Less reaction, more bet-placing. Less describing what you’d like, more shipping the thing that says what’s next.¶

The taste differential still holds. The window is shorter than I wrote last week. But the deeper differential — production-side participation, contextual application, forward-pointing conviction — isn’t on the same clock. That’s the layer the memory file doesn’t reach.¶

The model will eventually know what you liked. It will never know what you’re about to make.¶