Knowledge Without Expiry Dates

Three years ago I wrote a note that Jonas was at Stripe, probably a good contact for payment work. Jonas has since left Stripe, joined a fintech startup, left that, and is now freelancing. The note doesn't say any of this. It's quietly, confidently wrong — and I had no idea, because the last time I looked at it was the day I wrote it.

This is the problem with personal knowledge management that nobody wants to talk about: the notes you're not looking at.

All the advice about PKM is about capture. The right structure, the right tags, the inbox-zero for your knowledge base. But there's a second problem that gets almost no airtime, and it's in some ways more dangerous than not having a system at all. Stale information isn't neutral. It's actively misleading.

What the psychologists found

In 1994, Hollyn Johnson and Colin Seifert published a study on what they called the "continued influence effect." Subjects were given information, told it was wrong, and then asked questions that drew on the original (incorrect) account. Even after the correction, subjects continued making inferences based on the misinformation — often at the same rate as subjects who had never received a correction at all. The wrong belief persisted even when they knew it was wrong.

Their conclusion wasn't that people are bad at updating beliefs. It's that misinformation, once encoded, integrates into the reasoning structures we use to make sense of things. Pulling it out doesn't just require knowing it's wrong. You need to replace the explanatory structure it was part of.

I read this and immediately thought about my notes. Most of my old entries haven't been "corrected" — there's been no retraction, no update, nothing to trigger a revision. They're just sitting there, written in my own hand, carrying an implicit certification of accuracy that has nothing to do with whether they're still true.

What librarians figured out

Libraries have been dealing with this problem for longer than personal computers have existed. The field has a word for it: weeding. The standard professional framework — from something called the CREW Manual, widely used in American library science — uses an acronym to guide which books to remove: MUSTIE. Misleading. Ugly. Superseded. Trivial. Irrelevant. Elsewhere available.

Misleading comes first. Not "outdated." Not "wrong." Misleading — because the problem isn't the absence of accuracy, it's the presence of false authority. A 1987 nutrition book that confidently describes dietary fat as the primary cause of heart disease isn't just out of date. It's a liability. A library that keeps it on the shelf is actively providing worse service than a library with nothing on the topic.

The weeding metaphor is apt because gardens with no weeding don't stay the same. They become something different — something that crowds out what you actually want.

The accumulation one-way valve

PKM tools are almost entirely built around one direction of flow: in. There's usually a quick-capture shortcut, a clip button, an integration with your read-later queue. There is almost never anything equivalent to MUSTIE — no mechanism for evaluating whether existing entries are still trustworthy, no visible decay signal, no scheduled review that treats stale information as a problem rather than just old context.

Samuel Arbesman's 2012 book The Half-Life of Facts documented something counterintuitive: knowledge in different domains becomes outdated at measurable rates. Physics papers lose about half their citation influence in 10 years. Urology literature decays faster. What drives this isn't the data going anywhere — it's that newer, better-supported information renders older conclusions misleading rather than just incomplete.

I don't know the half-life of a note about a colleague's employer. It's probably somewhere between one and three years. What I know is that my knowledge base doesn't have any representation of this. Every entry carries the same implicit weight: present-tense, authoritative, true.

What structure makes possible

Here's an argument for typed, structured knowledge that I don't hear made often. The usual argument is that structured data is easier for AI to query — more reliable retrieval, less hallucination. True enough. But there's a more fundamental point: structure creates the precondition for systematic weeding.

A person record with a role field and a last_updated timestamp is something you can evaluate against external reality. You can ask which person records haven't been confirmed in over a year, which preferences predate a significant life event. Unstructured prose can't be weeded this way — there's no discrete unit to evaluate, no age attribute to sort on, no type to filter by.

Without that structure, what you have is an archive. And archives, unlike knowledge bases, aren't designed to be trusted. They're designed to be preserved.

The distinction matters more than it might seem. I want a knowledge base — something accurate in a current-tense sense, something an AI assistant can reason from without silently drawing on a note about Jonas that was true in 2021 and hasn't been since. That requires something libraries have had for decades: a policy about what to keep, and a mechanism for finding what should probably go.

---

Asgeir Albretsen is the founder of Harbor.