The one thing it couldn't look up

I tried to measure whether my own notes make a local model smarter. The first run said no. The reason was the twist: my always-loaded instructions were already feeding it the same rules, so there was no gap left to fill. The cortex moved the needle on exactly the one answer nothing else could supply.

I wanted a number for something I had been assuming: that feeding a local model my own accumulated notes makes it give better, more me-specific answers. So I built the obvious test. Ask the model a question whose right answer lives in my notes, once with the notes and once without, and count the difference.

The first run said the notes did nothing. Same answers with or without them. I nearly shipped that as the finding: personal context, no measurable effect.

Then I read what the “without” condition had actually written. It was recommending my own internal tools by name, citing rules that exist only because I wrote them down. That is not a model running blind. The bare baseline was not bare at all. The instruction file every session loads, the one sitting at the top of my setup, already carried those rules, so the model had them the whole time. I had been measuring my notes against a copy of my notes.

So I took that away and asked the raw model directly, nothing pre-loaded, the only difference being the single note I pasted in. Now the picture came clean. On almost everything, the note still changed nothing, because the answer was something any competent model reasons out on its own: keep private data on a local model, track a durable task as an issue, use conventional commits. General good sense. No gap to fill.

On exactly one question it flipped completely. Which tool to use for automatic versioning. Bare, the model shrugged or guessed the industry default. With the note, it named the specific custom action I actually use and told me to avoid the standard ones, because that choice is not derivable from anything. It is mine. It was the one thing it could not look up or reason its way to.

That is the whole shape of it, and it matched a separate test on injected skills exactly: augmentation helps only where it fills a real gap. The uncomfortable corollary is how narrow that region is. Most of what you would lovingly store and re-feed, a decent model already has, or can work out. The value concentrates in the handful of things that are true for you and almost no one else. Build the memory for those. The rest, it already knows.