Notes are written by your remembering self

Read a note you wrote two years ago about a decision. It probably sounds certain. Not the decision you remember making — the one that took three weeks and almost went a different way — but a clean version: "Decided to use X. Solves the problem cleanly." That was it. What the note didn't contain: the serious consideration of option Y. The week you were nearly convinced by Z. The specific, half-forgotten thing that finally tipped you toward X. The genuine doubt, held for longer than the note suggests was possible.

The note said "decided." What actually happened was slower and messier and more honest than that.

Two selves, one note

In 2011, Daniel Kahneman formalized something psychologists had been documenting for years: we have two selves that evaluate experience differently. The experiencing self lives in the present — it registers confusion, uncertainty, the back-and-forth. The remembering self constructs a narrative afterward, and that narrative doesn't average the experience. It uses what Kahneman called the peak-end rule: it remembers the most intense moment and how things ended, and discards most of the rest.

In a 1993 study, participants submerged a hand in 14°C water for sixty seconds — genuinely uncomfortable. In a second trial, they did the same sixty seconds, then kept the hand in for thirty more as the temperature rose slightly to 15°C. Still cold, still unpleasant, but fractionally less so at the end. A majority preferred to repeat the second trial. More total pain, remembered better because it ended better.

The experiencing self accumulates everything. The remembering self uses peaks and endings, and notes are written by the remembering self.

What the record actually contains

When you write a note about a decision, you almost always write it after the decision is made. You write the conclusion, because the conclusion is what you have. The experiencing self — confused, weighing options, not quite there yet — has no notes. It's too busy living.

So your knowledge base has a systematic shape, and the shape isn't random. It's missing exactly the information that would make it most useful: the alternatives you seriously considered before rejecting them, the things you weren't sure about at the time, the context that made one option feel right given constraints you've since forgotten. What it contains, mostly, is resolved conclusions written with a confidence that was usually retrospective. The note says "decided." The reality was "eventually came to believe, with reservations."

This isn't a failure of discipline. It's structural. The experiencing self can't write — it's still living through the thing.

The AI reads the highlight reel

This matters more now, because AI reads your notes.

When an AI agent queries your knowledge base, it's reading the remembering self's version of your history. It finds confident conclusions, curated summaries, decisions with stated reasons. It has no access to the uncertainty that preceded them. It cannot see the parts that didn't make it into the note.

The AI treats your knowledge base as a faithful record. It isn't. It's a selection — shaped by what felt worth capturing at the moment of resolution, written from the vantage point of having figured something out, edited unconsciously toward coherence. What feels like a reliable knowledge store is closer to a retrospective narrative, with the rough parts smoothed out. An AI reading it will be confident in ways your thinking never was.

This isn't a criticism of AI. It's a structural problem with what we hand it.

What the form can do

Michael Nygard's Architecture Decision Records — published in 2011, now standard in many software teams — solve a version of this problem. An ADR has fixed sections: context, options considered, decision, consequences. You fill it in before you close the question. The format forces you to document the uncertainty while you're still in it.

ADRs work not because engineers became more disciplined, but because the form makes a retrospective highlight nearly impossible. "Options considered" is either filled in or it isn't. The blank field exerts pressure.

Typed entities in a knowledge base work the same way. A decision record with an "alternatives considered" field produces different behavior than a freeform note. A person record with a "confidence" or "last updated" field captures different information than a scattered mention in a doc. Not because of discipline — because the schema requires you to notice what you don't know.

The experiencing self never makes it fully into the knowledge base. That's probably fine; we can't capture everything. But structure can at least hold space for it — a labeled slot where the uncertainty was supposed to go, a field that reminds you that conclusions had reasons, that decisions had alternatives, that confidence was earned rather than assumed.

The remembering self will write the note anyway. The form can at least ask it harder questions.

---

Asgeir Albretsen is the founder of Harbor.