What the Archive Still Believes
Notes capture what you thought. Nothing in a notes app knows when you changed your mind — and an AI reading that archive inherits every old belief.
Four years of notes, and the AI surfaced something I'd written about a former colleague in 2021. That he was unreliable. That I'd have to manage around him. I don't believe that anymore. We worked closely together for a year, and I was wrong about him. But I never updated the note — why would I have? The relationship had changed, the moment had passed, and retroactively correcting old notes is not a habit anyone naturally builds.
The AI didn't know any of that. It read the entry and factored it into its picture of the person. It was, technically, being diligent: using available evidence. The evidence was just a photograph of a view I'd since discarded.
This is the belief revision problem. It's older than software.
What philosophers figured out in 1985
In 1985, Carlos Alchourrón, Peter Gärdenfors, and David Makinson published a paper on the logic of theory change. It's been cited more than four thousand times, which is unusual for work that dry. The paper asked a precise question: when a rational agent encounters new information that contradicts a prior belief, how should they update their knowledge base?
Their answer involved eight formal postulates, but the central one is minimal change: revise only what you must to accommodate the new belief, preserving as much prior knowledge as possible while ensuring consistency. The goal is a knowledge base that doesn't contradict itself, updated with minimal disruption to what was already true.
No notes app I've used implements any of this. They support one operation: expansion. Add a note. The archive grows. There is no mechanism for revision — no way to mark an old note superseded, no structure for saying that the belief it expressed has been retracted. Alchourrón and colleagues were writing about formal logic, but they were also, without knowing it, describing the design gap at the center of personal knowledge management.
What medicine figured out earlier
Medical records took a different path, out of necessity. When a clinician makes an entry and later discovers it was wrong, they cannot delete it. The CMS guidelines are explicit: draw a single line through the erroneous content, keeping the original legible, then document the correct information alongside it with the current date, time, and reason for the correction.
This isn't bureaucratic caution. It's an answer to an epistemological problem. Medical decisions get made by people who weren't in the room when the original entry was written. They need to know not just what was recorded but whether it's still accurate — and if not, when it changed and why. The chain of amendments is itself evidence. An entry without that history, displayed as current fact, is dangerous.
Most personal knowledge tools have no concept of this. An old note looks identical to a current one. There's no visual indicator that the belief it contains was formed in a different context, nothing that ages into irrelevance. The archive is a flat layer: everything equally authoritative, everything equally accessible to an AI that cannot distinguish a live belief from a five-year-old one that you've since reversed.
Why typed entities narrow the problem
This is where structure buys you something real. In a person record with defined fields — employer, relationship notes, last contact date — updating one field is a revision, not an addition. The previous value is replaced (or preserved in version history); the new value is current. When an AI queries the record, it gets a consistent, single picture rather than a pile of potentially contradictory notes sorted by recency.
The same applies to decision records and preference records. A preference stored as a typed entity with a review date carries implicit metadata that a paragraph in a document doesn't: this was accurate as of a specific time and should be reconsidered on a schedule. Not a guarantee of accuracy, but a structure that makes staleness detectable in a way that unstructured prose never is.
The honest version of this argument is that typed entities narrow the problem without solving it. A person record can be updated cleanly; a document that draws inferences from a fact that no longer holds sits wrong in a way that's almost invisible. The cascade of a belief revision — all the downstream places that were built on the now-revised fact — is still mostly untracked. A typed entity is a named anchor; the prose around it is still opaque to change.
Leon Festinger noticed in 1957 that people actively avoid noticing inconsistency. We rationalize, look away, don't go looking for the places where our beliefs contradict each other. That's a reasonable strategy for getting through the day. It is a structural problem in a knowledge base that you're handing to an AI that will read every entry with equal attention, no protective inattention, no understanding that some things were provisional or have since expired.
The amendment you didn't write
Nobody is going to maintain perfect consistency across years of notes. The effort would make you stop writing anything worth keeping. That's not the goal.
The goal is narrower: structures that make staleness detectable, and operations that make correction explicit rather than invisible. What medical records got right is not that they were always accurate — it's that amendments are visible, dated, and reasoned. You can reconstruct not just what was believed but when it changed and why. The archive is legible not just as content but as a history of revision.
Most personal knowledge bases have no theory of belief revision. They accumulate. They were designed for input, not for the full lifecycle of a belief: formation, refinement, revision, retraction. Alchourrón and his colleagues formalized what rational belief revision requires in 1985. It took four thousand citations for that idea to show up in formal systems. It still hasn't really shown up in the tools most people use to manage what they know.
The single strikethrough, kept legible, dated, with a reason. Forty years before the AGM postulates, before SQLite, before anyone thought about giving an AI access to your notes. Still the clearest answer to the problem.
Asgeir Albretsen is the founder of Harbor.