Notes Are Not Written to Be Read

In 1978, Norman Slamecka and Peter Graf ran a deceptively simple experiment. They gave subjects word pairs — "KING-CR___" — and asked them to complete the blank. Then they tested recall. People who generated the word themselves remembered it significantly better than people who simply read the completed pair "KING-CROWN." The effect held across dozens of variations — cued recall, free recall, recognition — and eventually got a name: the generation effect.

The implication for note-taking is uncomfortable. Writing doesn't primarily store information. Writing creates memory. The note is a byproduct — cognitive scaffolding that served its purpose in the moment of being built. Slamecka and Graf's finding wasn't specifically about notes, but every note-taking researcher since has pointed at the same thing: the cognitive work happens during writing, not after.

This is why researchers have found that notes you never re-open often aren't worth what they cost. You spent the cognitive load of capture and skipped the gains of review. The writing helped you think. The artifact just sat there, doing nothing.

That math holds for humans. It doesn't hold for AI agents.

The silent reader

The first time an agent reads through your notes looking for context, it encounters something its designers never quite planned for: an archive of documents written for the act of writing, not for reading.

When I write a note after a difficult meeting, I'm not writing with full fidelity for future-me. I'm externalizing what I need to hold while I process everything else. Fragments. Half-sentences. References that assume the whole situation is still loaded in working memory. "Talk to Elena about the timeline" — useful tomorrow. Unresolvable six months later, when I can't remember which project, which Elena, or whether that conversation ever happened.

These notes aren't broken. They worked exactly as intended. The generation effect fired. I remembered what I needed to.

But an agent has no memory of writing them. It reads them cold.

What the agent doesn't know

There's a failure mode here that's different from notes going stale over time. Stale notes are a problem of past-you failing to write for future-you. This is something else.

A note written mid-thought is dense with implicit reference — pronouns and phrases that only resolve if you share the frame. "It," "the decision we made," "the thing she mentioned." In conversation, these work because everyone is present. In a cold retrieval, there's no shared frame. The text is there. The meaning isn't.

Most knowledge bases are also heavy on exactly the kinds of notes most vulnerable to this. Meeting notes. Quick capture. Journaling. These are the high-volume, low-structure formats that note-taking apps optimize for — the easiest to write and the worst to retrieve from. An agent that treats this archive as a reliable source of personal context will be confidently wrong in ways that are hard to detect. Not hallucination — something more insidious. It'll surface real notes that carried real meaning once, read them as statements about the present, and act on them.

And it won't flag its uncertainty the way a human might. A human reading your old notes knows they're reading old notes. The agent doesn't experience that degradation.

The problem Luhmann solved

Luhmann built his Zettelkasten differently. Each slip was written as if for a reader — explicit, self-contained, linkable, with no assumed context beyond what was on the card. In his 1981 paper "Communicating with Slip Boxes," he described the system as a "communication partner": something that could respond to him later with ideas he'd forgotten he had. That required a specific discipline — writing for retrieval, not just for encoding.

Most people don't do that. Most notes apps don't push you toward it, either.

What was built to be read

The knowledge that holds up under agent retrieval is structured. A person record that includes enough context to interpret without ambient memory. A decision that captures why, not just what. A preference stated explicitly instead of implied through surrounding prose.

These aren't written just to think. They're written to persist.

The generation effect still applies — writing a decision record probably does help you think through the decision more clearly than just deciding. But the artifact you produce is designed for a reader who wasn't there. The pronouns resolve. You can hand it to an agent and get back something coherent.

Notes are acts of thinking. Structured records are acts of documentation. Most knowledge bases conflate the two, which works fine when only a human reads them — a human fills in the gaps from memory. An agent can't. It only has the text.

Whether that's a design flaw or just an honest description of what notes have always been, I'm not sure. Maybe both.

---

Asgeir Albretsen is the founder of Harbor.