The Search You Never Ran
Search assumes you know what you're looking for. The more interesting problem is surfacing what you didn't know you needed.
You're halfway through writing a proposal when you realize you've made this exact argument before. Not exactly — different client, different context — but the same core case. The same objection you'll need to address. The same framing that worked last time. You're sure you wrote it down somewhere.
You don't find it. You write the proposal from scratch, do fine, move on. Three weeks later, cleaning out old documents, you find the original note. It was there the whole time.
This happens more than people admit. Not because they failed to capture the information — they did. Not because the search is broken — the search works fine, if you remember to run it. The problem is that you didn't know to look.
What search assumes
Every search tool is built on the same assumption: you know what you're missing. You type "client proposal objection" and the system finds documents containing those words. The architecture is reactive by design. It waits.
This is fine for most things. If you need to find a specific person's contact details, search works exactly as well as you need it to. But a lot of the most valuable knowledge isn't retrieval of specific facts — it's recognition of connection. The realization that the problem you're solving today is structurally similar to something you navigated two years ago. The note about a colleague that turns out to be relevant to a completely different project. The decision you documented that quietly constrains what you're planning now.
For that kind of knowledge, search is the wrong interface. You'd have to know what to search for before you could find it. You can't search for what you don't know you've forgotten.
What Bush actually described
In July 1945, Vannevar Bush published "As We May Think" in The Atlantic. It's remembered mostly for predicting hypertext and the internet, but the thing Bush actually cared about was different. He called it associative indexing.
The memex — Bush's hypothetical desk-sized device — would store your documents, your notes, your photographs, your correspondence. But the key feature wasn't storage or even retrieval. It was trails. "Any item may be caused at will to select immediately and automatically another," he wrote. You'd build chains of association as you worked, and the memex would follow them. Working on topic A would surface related items you'd linked to years earlier when working on something else entirely.
The web borrowed Bush's hyperlinks. It didn't borrow the part where the links are yours.
On the web, links are embedded by the publisher. You can't add your own association between two pages you found relevant. You can't leave a trail through your own reading that the system will surface next time you're near the same territory. What Bush described was a memory expansion device. What the web built was a publishing platform.
What changes with AI
The honest answer: not everything, and not yet. Most AI-assisted knowledge tools are still reactive. They've upgraded the search box — semantic search instead of keyword matching, natural language questions instead of query syntax — but the underlying model is the same. You ask; the system answers.
What actually changes is that AI can now read your entire knowledge base contextually and notice things you wouldn't have thought to ask about. Working through your notes on a new project, it might recognize that you've already documented a relevant constraint — in a different folder, from a different year — without you mentioning it. Not because it searched for a keyword. Because it read everything and made the connection.
This is closer to what Bush was describing. Not smarter search. Associative surfacing.
The limitation is real: relevance detection without surveillance is a hard problem. A system that proactively surfaces everything potentially relevant becomes noise very quickly. The signal-to-noise ratio depends entirely on how well the system understands what you're actually working on — and what actually counts as relevant to you, not just textually similar.
Structured knowledge helps here in ways that raw text doesn't. If your knowledge base treats people, projects, and topics as typed entities — not just words in a document, but records with relationships — then "you're working on Project X, which involves Person Y, who also appears in a note you wrote eighteen months ago about a similar situation" becomes a tractable query. The system isn't pattern-matching on keywords. It's traversing a graph of things you've said matter to each other.
The shape of a useful suggestion
The one thing I've noticed about genuinely useful proactive suggestions is that they're specific enough to be actionable and unobtrusive enough not to interrupt. The worst version is the notification model: the system pings you with "did you know you wrote about this?" The better version is ambient — when you open a document or start a conversation, relevant context is quietly available without demanding attention. You notice it when you need it. You don't when you don't.
Bush imagined people building trails as a matter of habit. Trails that a researcher might bequeath to their students alongside their notes. The trails were the intellectual work, not just the documents themselves.
Most of us don't build trails. We write things down and hope we'll remember to search for them later. AI doesn't change the writing. But it might finally change the finding.
Asgeir Albretsen is the founder of Harbor.