Memory.

Every correction you've ever given a coding agent was a piece of the spec. "No, sans-serif." "The avatar opens a hovercard, it doesn't navigate." "We use the outbox pattern — I told you this three weeks ago." You said each one exactly once, which should have been enough, and then it scrolled away.

That last quote is the tell. If you've ever said "I told you this already" to a machine, you've located the actual failure — and it isn't intelligence. The model that got it wrong today is smarter than the one that got it right in March.

Every repeated correction is a failure of memory, not intelligence.

So the golden rule I build by: each time the human has to step in, the system must learn so the human never has to say the same thing again. One correction, permanent. Anything less turns you into a tape loop, re-recording direction you already gave — and your direction is the most expensive signal in the entire system.

Which is why summaries make me nervous. A summary is lossy compression with no error term — nothing tells you what got dropped, and what got dropped is usually the nuance you fought for. Keep receipts instead: the actual words, cited to the actual moment they were said.

And when I contradict myself — I do, constantly — that's not noise to clean up. If a thing worked as A and now works as B, the seam between A and B carries information the current state can't: why A seemed right, what broke, what B is protecting against. A system that stores only present truth has amnesia about its own reasons. Truth is a trajectory, not a point.

There's a simple test for whether any of this works. After months of giving a system direction — taste, vision, corrections, reversals — can it take one instruction and produce exactly what you meant? That's the north star. Not smarter models.

A system that was listening the whole time.