The Chapter Your Knowledge Base Doesn't Know About
When a chapter of your life ends, your notes don't know. The problem isn't just stale facts — it's entire context frames becoming invalid at once.
You ask an AI to help you draft a quick email to someone you used to work with. A warm reconnect, a low-stakes note. The AI surfaces what it knows about her: she's your team lead, prefers written agendas, gets defensive about scope changes, books calls on Tuesdays. It calibrates your draft around this. The email comes out careful, hedged, a little deferential.
You left that job eighteen months ago. This isn't a work email. It's a casual catch-up. The AI just handed you an interaction script for a version of the relationship that ended with your last day.
This will happen more as knowledge bases get AI-connected, and it happens for a structural reason that has nothing to do with bad engineering.
The chapter as a unit of obsolescence
Dan McAdams, a psychologist at Northwestern, has spent decades studying how people construct identity through narrative. One of his findings is that we naturally segment experience into chapters: periods of life organized around a context. A job. A city. A relationship. A project. When that context ends, the chapter closes.
The knowledge you accumulated during a chapter was calibrated to that chapter. Your notes about a colleague's communication style made sense in the context of a shared team with a shared goal. Your preferences captured during a period of heavy client travel reflect a version of you who no longer works those hours. Your financial notes from 2022 were written in an inflation environment that no longer obtains.
Most personal knowledge tools are built as additive systems. Notes accumulate. The assumption, usually implicit, is that newer notes are more relevant, but older notes remain valid. That's true for some things: your date of birth, your medical history, the phone number you've had for ten years. These facts don't have chapters. But most personal knowledge does.
The failure isn't any individual note going stale. It's the entire context frame expiring at once.
Stale knowledge is worse than no knowledge
Kevin Munger, a political scientist at Penn State, introduced a concept he calls "knowledge decay" in a 2019 paper. His argument: knowledge doesn't just become irrelevant as the world changes. It becomes actively misleading, because it was calibrated to a world that no longer exists. An email drafted with chapter-stale context doesn't give you neutral errors. It gives you confident, calibrated ones.
This is a meaningful distinction. An AI that knows nothing about your former colleague writes a generic email. An AI that knows a lot about the wrong version of her writes a confidently wrong one.
The harder problem is that you usually can't tell. The draft sounds right. It has the right tone, the right details. The calibration is subtle. You might send it and wonder later why the response felt off.
Temporal knowledge graphs solve this for public facts. Systems like Wikidata model every claim with explicit validity windows: "Jane held the role of CTO from 2018 to 2023." The fact doesn't disappear when her tenure ends; it gets timestamped and archived. Queries against current state return current facts; queries against history return historical ones.
Personal knowledge tools have not imported this pattern. They mostly have no concept of "was true then, not now."
What typed entities make possible
The note format makes this hard to fix. A prose note written during a project is densely embedded in the context of that project. There's no clean seam between "information about the person" and "information about our relationship at that time." When the project ends, you can't surgically extract the durable facts from the contextual ones.
Typed entities can do this. A person record with explicit fields can hold a "relationship context" that notes where the relationship began and what it was. "Former colleague at Acme, engineering team, 2021–2023" is a retrievable fact. It tells an AI reader: this was the frame for what follows.
Project records work the same way. A project with a status of "shipped" or "archived" signals that the context was a chapter. Notes under it are history, not current guidance.
This isn't about deletion. Old chapters are worth keeping: the history of your thinking, the record of decisions and relationships, the context you might need when a relationship resurfaces years later. But there's a real difference between an archive and active context. An archive is true-in-the-past. Active context is what shapes behavior now.
Most knowledge bases don't have a word for that difference.
The map that wasn't updated
When knowledge accumulates in a tool that doesn't understand temporal context, the knowledge base slowly becomes unreliable in a specific way. Not wrong, exactly, but set in a world that partially no longer exists. Like a map that was accurate when printed and hasn't been updated since.
An AI reading this map won't know which roads have changed. It will navigate with the confidence of a current map. It will recommend turns that no longer go anywhere.
Chapters are how a life is actually organized. Not as a continuous feed of present-tense facts, but as periods with beginnings and ends, each generating knowledge that was meaningful in its time. A knowledge base that can't represent this will eventually be full of things that were true and treat them as things that are.
Asgeir Albretsen is the founder of Harbor.