Two Kinds of Memory, One Kind of Notes App

You want to know whether John prefers morning or afternoon meetings. You search your notes. You find: a note from June 2023 about a call that started late and ran over. An email thread about a proposal. A brief from the kickoff where someone mentioned timezone conflicts. Somewhere in those fragments, the answer lives. But you can't retrieve it — only excavate it.

The problem isn't that you didn't write it down. You did. The problem is that you wrote it into an event, not as a fact.

Two kinds of memory

In 1972, Endel Tulving published a chapter that would reshape cognitive psychology. His argument: long-term memory isn't one system. It's at least two. Episodic memory holds events — specific experiences anchored in place and time. Semantic memory holds knowledge — facts, concepts, relationships, stripped of context. You remember finishing a book on a particular flight (episodic). You know the book was by Borges (semantic).

The distinction sounds academic until you sit with it. Tulving pointed to patients with amnesia who couldn't form new episodic memories but retained semantic knowledge: they couldn't tell you what they'd done that morning, but they still knew words, recognized faces, held facts about the world. The two systems could be damaged independently. That independence is the point.

Your brain runs both systems simultaneously and invisibly. When you read a meeting note from 2022, you extract the semantic content — "right, Sarah doesn't like being surprised in reviews" — and file it away without thinking. The event fades; the fact stays. This extraction happens automatically, in the background, and costs you nothing.

Notes apps do none of this. They store the events. The extraction is left entirely to you.

The note is always an episode

Almost every note has an implicit "when" encoded into it: after the meeting, during the call, this morning, while reading. A note is a trace of an event. The knowledge you wanted to capture is somewhere in the event, but it isn't the note's structure; it's an inference the reader has to make.

A note that says "talked to Marcus today — he mentioned he's switching teams in the fall, new manager is Torres" is an episodic record. The semantic facts embedded in it are: Marcus is changing teams. His manager is Torres. Both were true at the time of writing. Neither one is the note.

When you search your notes later, you're searching through events. If you remember roughly when something happened, or a keyword from the note, you find it. If you're just looking for "who manages Marcus now?" — you're in trouble. The fact doesn't live anywhere in particular. The event does.

This isn't a problem that better search solves. Full-text search and semantic vector search both operate on what's actually in the notes. If the note is "talked to Marcus today," neither approach will reliably answer "who manages Marcus" unless that document surfaces in context. It might. It often won't.

The AI makes it worse before it makes it better

When you connect an AI to your notes, you're handing it an archive of events and asking it to reason about facts. The AI doesn't have your brain. It can't run the automatic extraction you perform when you re-read something. It reads "talked to Marcus today — he mentioned he's switching teams, new manager is Torres" and absorbs the information — but only if that specific note surfaces in retrieval. If the note is old enough, or buried, or the query doesn't trigger it, the AI operates without it.

Worse: the AI can't flag what it doesn't know. If your notes contain no person record for Marcus, no explicit statement of his current manager, the AI will either hallucinate an answer or report uncertainty. The episodic archive doesn't help it; it helps the version of you who already knows what to look for.

Tulving's distinction has a direct architectural implication. If you want AI to reason about facts, the facts need to live somewhere explicit. Not embedded in events. Not accessible only to a brain that can extract them. Somewhere typed and retrievable: a person record that says "Marcus / team: growth / manager: Torres / as of: fall 2023."

The extraction problem

The reason most knowledge bases don't have this layer is that extraction is work. It's faster to paste a note and move on than to decide: what kind of thing is this? A preference? A person fact? A decision with consequences? Capturing an event takes seconds. Converting it into semantic knowledge requires judgment — about what the fact actually is, how confident you are, what context it might need later.

But this is the moment that matters. At the point of writing, you have all the context. You know what the fact is, where it came from, whether it's reliable. A week later, you have an event log. A month later, you have a keyword. A year later, you have nothing you can act on.

Typed entities — person records, preference records, decision records — are what happens when you do the extraction at capture time, while you can still do it reliably. They don't replace notes. They're what notes were always trying to become.

Tulving's amnesiac patients couldn't form new episodic memories. They still knew things. The semantic system ran independently, on whatever had been properly encoded as fact rather than as experience.

Most of your notes will survive as events. The ones worth something when you actually need them are the ones someone converted into facts — past-you, at the moment of writing, before the context dissolved.

---

Asgeir Albretsen is the founder of Harbor.