The Stranger Who Left You Notes
Your knowledge base is a collaboration between every version of you. The one who wrote the notes might not agree with the person reading them.
A while back, my AI recommended a sushi restaurant I genuinely dislike. I traced the suggestion back through the conversation and found its source: a preference note I'd written in early 2021, logged with confidence, sitting in my knowledge base as current fact. I'd liked that place once, apparently. Then something shifted. The restaurant, or me, or just the memory of it. I never updated the record.
The note wasn't wrong. It described a real preference. It just described a preference I no longer have.
This is easy to dismiss as a data quality problem: stale records, lazy maintenance, a missed update. But I've been sitting with it, and I think it's something more structurally strange. The data was fresh when I wrote it. It wasn't inaccurate at write time. The problem isn't that the note changed. The problem is that I changed, and the note didn't know to.
The degree of it
In 1984, the philosopher Derek Parfit argued in Reasons and Persons that personal identity is a matter of degree, not kind. You don't persist through time as a fixed entity. You exist as overlapping chains of psychological connection: memory, character, beliefs, values, tastes — each linking you to your recent past, which links to its past, and so on. The connection is strong between yesterday and today. It weakens across years. The version of you who wrote something in 2021 is not a different person, exactly. But she's not fully the same person either.
Parfit was doing philosophy, not product design. But the implication is precise: the author of an old note in your knowledge base is, in a meaningful sense, not quite you. The difference is a matter of degree. Sometimes it's trivial. Sometimes it's the sushi restaurant.
Research by Rodica Damian followed 1,795 people across 50 years, from age 16 to 66. Average trait-level change across the Big Five personality dimensions was half a standard deviation. Between 20% and 60% of participants showed reliable change within any given trait. Damian described some of the shifts as "very, very large" — clearly visible to outside observers. Whatever stability we have doesn't come from staying the same. It comes from changing slowly enough that we rarely notice.
The confidence problem
In 2013, Jordi Quoidbach, Daniel Gilbert, and Timothy Wilson published a study of more than 19,000 people across all age groups. Each person was asked how much they'd changed in the past decade and how much they expected to change in the next. The finding was consistent at every age: everyone believed they'd changed a lot in the past and would change very little from now on. Each version of us experiences the present as a stable endpoint. They called this the "end of history illusion."
It's a precise description of how preferences end up in knowledge bases. You write a note about what you like, what you want, what you value — with complete sincerity, in the quiet belief that this is just you. Not a snapshot of a passing state. Just you. The 2021 version of me wrote that preference note with the same confidence I'd apply to something I believe right now. And the statistical expectation is that I'm currently writing records that will, at some point, describe a person I used to be.
This is what makes it a design problem, not just a cleanup problem. Even diligent maintenance — regularly reviewing, updating, deleting — only addresses records you know to revisit. It doesn't address the underlying fact that every note in your knowledge base was written by a slightly different author who believed they were writing timeless truths.
When an AI reads your knowledge base, it reads it flat. A preference from 2021 and one from last week sit with equal authority. The AI has no way to know the older one was written by a version of you with different work, different relationships, different ideas about what a good evening looked like. The records don't carry that.
What structure can do about it
Free prose can't be interrogated for staleness. A paragraph doesn't have a schema. You can't ask a blob of text how out of date it is.
Typed entities can at least be asked. A preference record with a last_reviewed date and a staleness threshold can be flagged before an AI cites it as current. A person record with a status field can be distinguished from an active relationship. A decision record with an outcome field distinguishes resolved choices from open ones. Structure doesn't prevent drift. But it makes drift detectable, which is the precondition for doing anything about it.
So a knowledge base with typed entities and review dates gets you a list of things to look at. That's not nothing. But the harder part comes after detection. You open the record. You read a preference that was true once. And you know — having read this — that your current preferences will eventually look like that to your future self. Not wrong when written. Just slowly, quietly, no longer yours.
The review date surfaces the record. What you do with it is still up to you.
Asgeir Albretsen is the founder of Harbor.