Why Your Notes Remember Everything Except Why You Wrote Them

You open a note you wrote sometime last winter. The text is perfectly legible — a short paragraph, a few bullet points, clean and purposeful. The idea inside it is sensible. But you can't remember why it's there. Was it a reaction to something? The seed of a decision you shelved? A thing you half-believed and then quietly dropped? The note survived intact. Its reason didn't.

This is the provenance problem. And it's not a failure of memory — it's a structural feature of how notes work.

Source is not a tag

In 1993, Marcia Johnson and her colleagues at Yale published what became one of the foundational frameworks in memory research: source monitoring theory. The core insight is subtle but consequential. When we form a memory, we don't attach a discrete label that says "I learned this from Dave on Tuesday." Source information is recovered later, at retrieval time, from contextual cues that were stored alongside the memory — emotional valence, perceptual detail, what else was happening in the environment, the cognitive effort involved in processing it.

When those cues erode — and they do, consistently, especially for semantic content — source attribution fails completely. People misattribute where they heard things. They confuse what they imagined with what they actually did. They take credit for ideas they encountered somewhere. Not because they're careless, but because the cues that would disambiguate the source have decayed past the point of use.

Notes have the same architecture problem. The why is never written into the note. It lives in your head at writing time — you know what conversation prompted this, what project it's attached to, what you were trying to decide. The note is an artifact of that context, not a record of it. Months later, you're in a completely different state, the context has dissolved, and you're reading the conclusion of an argument whose premises you've forgotten.

What this actually costs

Most of the time, nothing. You read old notes, shrug, move on.

But the costs compound in two specific places. The first is decisions. Notes are often the traces of half-formed thinking — positions you were working toward, options you were weighing. Without provenance, it's easy to overweight them. An idea you wrote speculatively during a brainstorm ("what if we tried X?") can read, fourteen months later, like a settled position. You act on it as if you meant it firmly. You probably didn't.

The second place is AI interaction. When you feed a personal knowledge base to an AI, you're handing it not just information but implied confidence levels, implied temporal relevance, implied purpose. A decontextualized note that says "use SQLite for this" looks the same whether it was a firm architectural decision or a throwaway thought you abandoned two sprints later. The AI doesn't know which. It can't recover the provenance. So it treats the words at face value, and now you have a system nudging you toward a premise you dropped.

This is where most discussions of "AI memory" miss the point. The question isn't only what information an AI can access. It's how much metadata is attached to that information — when it was written, under what circumstances, and what status it carried at the time.

The metadata that matters

Luhmann understood some of this intuitively. His Zettelkasten system — nearly 90,000 index cards organized across four decades of work, now archived at Bielefeld University — wasn't just about linking ideas. Each card had a precise alphanumeric address, a temporal sequence, a referential chain back to whatever it was responding to. You could read a card and understand roughly when in the intellectual project it appeared, what it followed from, what it preceded. Not perfect provenance. But structurally richer context than most digital notes carry today.

Modern tools mostly handle this badly. A Notion page has a creation timestamp buried in its metadata, invisible during normal use. An Obsidian note has no inherent sense of what project prompted it. An AI memory that stores "user prefers async communication" has no record of whether that came from something the user said once in passing, or from a deliberate preference they explicitly set — and that difference matters enormously when you're deciding whether to act on it.

The better design treats provenance as a first-class property. Not something scraped from a creation date, but something captured intentionally: what conversation prompted this, what the status was (speculative, confirmed, abandoned), what triggered the most recent change. When AI tools create or update knowledge, logging why they did it is as important as logging what they changed. Diffs without context are only half the story.

Not quite a solution

I'm not going to suggest this is easy to get right. There's a fundamental asymmetry between writing time and reading time that no system fully resolves. You can't always know, in the moment, what you'll need to recover later. And there's a practical limit to how much metadata anyone will actually maintain before the overhead kills the habit.

But the framing matters. If you build notes as containers for conclusions, you'll capture text and ignore context. If you build them as snapshots of reasoning — with the temporal and situational scaffolding that reasoning involves — you'll build systems that preserve the conditions under which the thinking happened.

That's a harder problem. It's also a more honest description of what knowledge actually is.

---

Asgeir Albretsen is the founder of Harbor.