The diver's problem

In 1975, Alan Baddeley and Donald Godden ran an experiment with scuba divers at a Scottish harbor. They asked eighteen divers to memorize lists of words — some on dry land, some underwater — then recall them either in the same environment or the other one. Divers who learned words underwater and recalled them underwater remembered about 50% more than divers who learned underwater but had to recall on land.

Same words. Same brains. Different context.

This is context-dependent memory, and it's well-established in cognitive psychology. What makes it interesting is the precise mechanism. The memory didn't vanish when the divers surfaced. The retrieval cue did. Endel Tulving, who formalized the encoding specificity principle two years earlier in 1973, put it clearly: "what is stored determines what retrieval cues are effective." The environment at encoding becomes part of the memory trace. Change the environment, and the thread back to the memory is severed.

Most people assume forgetting is a storage problem. It's usually a cue problem.

The note-taking version

Every note you've ever written has a Godden-Baddeley problem embedded in it.

When you wrote the note — let's say a decision you made, or something a colleague mentioned, or a half-formed idea you wanted to come back to — you were in a specific context. A project was active. A particular goal was in frame. You'd been reading certain things for the past few weeks. You had a vocabulary for the moment. And you wrote the note in two minutes because the surrounding context made it legible without explanation. You knew what you meant.

Six months later, you've surfaced. The project is over or renamed or quietly shelved. You're thinking about different things. You try three different search queries and find nothing useful. You move on.

The note is there. The cue isn't.

The failure isn't in your search tool. It's not that your notes are poorly organized. It's that the search you're running now uses the vocabulary and associations of present-you, and the note was written in the vocabulary and associations of past-you, in a context that no longer exists. Tulving's principle applies across time, not just space. The retrieval context has to match the encoding context. It doesn't.

Full-text search makes this specific failure invisible. You can only query what you remember you were thinking about when you wrote the note. That's precisely the information you don't have.

What a type actually does

The usual argument for structured knowledge — typed entities, categories, schemas — runs through organization. Notes with types are easier to sort, browse, and query.

That's true. But it misses the more important thing.

A type is a context-independent retrieval handle. When a note about a person is stored as a person record, with a name as the primary key, the name is a stable cue that survives context collapse. It doesn't matter that you wrote it in the middle of a project that's since ended, using terminology you've drifted away from, in a mental frame you can no longer reconstruct. "This is about Jordan Kim" is a handle that still maps onto how you'll think about it years later.

The same is true for decisions, preferences, tasks. A decision record stays findable because the category "I made a decision about this" is stable. It corresponds to how future-you will approach the question, even if the specific vocabulary has drifted beyond recognition. A raw prose note about the same decision has no such stability. It lives in the language of its moment.

This is the most underappreciated argument for typed entities: not that they're more organized, but that the type acts as a cue that survives what the prose does not. The category doesn't care when it was written.

The honest limit

None of this fully solves the problem.

Cue-dependent failure happens inside structured systems too. A person record with a name, one email, and three sparse notes from 2022 will still be hard to surface at the right moment. The type gives you the first step of retrieval. What you need next is enough signal within the entity to make retrieval meaningful, not just possible.

And there are notes that genuinely resist typing. Not everything is a person, a task, or a preference. Some of what you need to retrieve is more like a train of thought — a half-formed argument that only makes sense if you can get back to the session that produced it. For those, the cue problem is harder, and no schema saves you.

But most knowledge isn't like that. Most of what people need to find later is findable, in principle. The note exists. The decision was recorded. The contact has an entry. The problem is the cue, not the content. And if the category acts as a cue that outlasts the encoding context, then it's doing something more than organizing your information. It's keeping the thread intact.

The divers remembered because the water was the same. The goal, in a knowledge base, is to give future-you something equally stable to grab onto — not the vocabulary of a specific moment, but a handle that doesn't change when the context does.

---

Asgeir Albretsen is the founder of Harbor.