Your Notes Have a Half-Life
Medical facts expire in 45 years. Engineering knowledge in a decade. Your personal notes have no such warning — and your AI agent won't notice either.
Six months ago, I asked an AI to help me draft an email to a former colleague. It pulled from my notes: his current role, the project we'd worked on, a few context points I'd written down after we last met. All of it sounded right. The job title was two roles out of date — he'd moved companies entirely, nearly two years before.
The note wasn't wrong when I wrote it. It just hadn't been updated. Nothing about it said check this first. It sat in my system with the same quiet authority as a note I'd written that morning.
I sent the email. He replied, clarifying, politely. The kind of correction that sticks with you.
The decay is already happening
In 2012, Samuel Arbesman published The Half-Life of Facts, which borrowed the metaphor of radioactive decay and applied it to knowledge. His central argument: facts don't last forever, and when you look at large bodies of knowledge, they expire at predictable, measurable rates. A team of researchers examined nearly five hundred articles about cirrhosis and hepatitis and found that half the medical knowledge in those papers became outdated within about 45 years. Engineering knowledge shrank faster: a century ago, half of what an engineer learned in school became obsolete within 35 years. By the 1960s, that had dropped to around a decade.
The point wasn't that knowledge is unreliable. The point was that knowledge expires — measurably, at different rates for different domains — and almost nothing in how we store information accounts for this.
Your personal knowledge base has the same problem. Worse, in some ways. Medical literature at least has retraction notices, updated guidelines, journal corrections. The mechanism is slow and imperfect, but it exists. When a finding is overturned, there's a path for that correction to propagate, however delayed. Your notes have no such path. The job title doesn't update itself. The project status you noted six months ago doesn't get revised when the project ships.
What unstructured prose can't tell you
The structural problem is that notes carry no metadata about their own reliability.
A plain text note — a Markdown file, a paragraph in Notion, an Obsidian page — has a creation timestamp and maybe a modification date. That's it. There's no field that says verified, no flag that says likely outdated, no schema that makes it possible to ask: which of these facts are more than two years old and haven't been touched?
When you're the only reader, you compensate for this implicitly. You remember, roughly, when you wrote something. You notice that a note mentions a role at a company you know the person left. You apply a kind of informal source monitoring — unconsciously, while reading — discounting older notes slightly, flagging things that feel uncertain.
An AI agent reading your knowledge base doesn't do any of that. It reads what's there. The two-year-old job title is a fact. The phone number you verified yesterday is a fact. Both are weighted equally, drawn on with equal confidence, dropped into the email it's about to help you write.
This isn't a hallucination problem. The AI isn't making something up. It's faithfully retrieving something you wrote, which is the most frustrating version of being wrong: it was you, just earlier.
The typed entity as a signal
The response isn't reviewing all your notes — that's a losing battle against accumulation. The response is writing a different kind of object.
A generic note has no schema. There's nothing to query that surfaces staleness. You can filter by creation date, but that's not the same as finding the notes where the facts are most likely to have drifted.
A typed entity is different. A person record has a current role field, a last contact date, associated notes with timestamps. When those fields are structured, you can query them. An AI reading a typed person record can note that the last-contact date is two years old — rather than assume the data is current. A preference record with a review date can surface in a review queue. A decision record with an outcome field can flag itself as needing closure.
This doesn't stop knowledge from expiring. It makes expiration visible, which is a different thing. Unstructured prose can't be incomplete in a detectable way. Typed entities can: the empty outcome field in a decision record isn't nothing. It's a gap that's visible, queryable, and therefore actionable.
Arbesman's radioactive decay analogy implies something uncomfortable: the decay is happening whether or not you're measuring it. Your notes are aging whether or not you check. The half-life of a person's job title — in an economy where people move often — is probably two or three years. The half-life of a project status might be six months. The half-life of a preference note is harder to estimate, but it isn't infinite.
The problem isn't that notes go stale. Every knowledge system has this problem. The problem is that stale notes look exactly like accurate ones. And when an AI agent reads your knowledge base, that invisibility becomes its problem too.
Asgeir Albretsen is the founder of Harbor.