Here is a habit worth noticing: almost every time you search for something, your query leaves the building. You type a few words, and those words travel to a server — to be matched, logged, ranked, and very often remembered. This is true of web search, obviously. But it is increasingly true even when you search your own stuff: your notes, your files, your messages. The query goes up to a cloud index and comes back with results. The thing you were looking for becomes a fact about you that lives on someone else's computer.

We wanted search that never does that. Not search that promises to be careful with your query — search where the query has no opportunity to be careless, because it never goes anywhere. Everything happens on your device. Your searches, your data, your results: none of it leaves your hardware. That is the whole premise, and the rest of this piece is about how it manages to be good search under that constraint instead of a worse one.

Two ways of matching, blended

Plain keyword search is fast and exact and a little bit dumb. Type "garden," and it finds documents that literally contain the word "garden." If the note you actually wanted said "backyard" or "growing space," it sits there invisible, because the letters do not match. Keyword search has no idea those words are about the same thing.

The fix is to also search by meaning. The system converts text into a mathematical representation that captures the concept behind the words rather than the words themselves — so "garden," "backyard," and "growing space" land near each other because they mean something similar, even though they share no letters. A search for one can surface notes that used the others.

Neither approach wins alone. Keyword search gives you precision when you know the exact term. Meaning-based search gives you understanding when you do not. So the system runs both and blends the results, taking the sharpness of exact matching and the reach of conceptual matching at the same time. And — this is the part that matters for the premise — both of these run locally. The model that understands meaning lives on your device. There is no remote step where your query gets sent off to be understood.

Search that follows connections

Finding matching documents is only half of what makes this useful. The deeper thing is that, as you use the system, it quietly builds a map of how your memories connect — a web linking people, projects, files, conversations, and ideas to each other.

So when you search, the system does not just hand back documents that matched. It follows connections. Search for a person, and you can surface not only their contact entry but the project you worked on together, the group you both belong to, the conversation where they mentioned the thing you are half-remembering. The match is a starting point; the connections are what turn a hit into an answer.

This map gets richer the longer you use the system. Every new note, every new conversation, every new link adds to it. After months of use you are not searching a flat pile of files — you are searching a connected picture of your own digital life, one that no one else can see, because it was assembled on your device from your data and never copied anywhere.

What it feels like

Let me make it concrete, because the abstract version undersells it.

You remember a conversation from months back. Someone recommended a book — something about forest gardening — and you want it now. But you cannot remember who said it, or when, or where the conversation happened. In a normal tool this is hopeless; you do not have a keyword to type, because you do not remember the words.

Here you can ask the question the way you actually remember it: who recommended that book about forest gardening? The system searches by meaning, not by exact words, so it finds the conversation even though you never had the precise phrase. Then it follows the connections — to the person who said it, to the group channel where it happened, to the book itself. It surfaces all of that together. The conversation was three months ago, in a particular channel, and it links straight to the contact who recommended it and the title they named.

And every step of that happened on your device, in seconds. The question "who recommended that book about forest gardening" — which is, if you think about it, a fairly intimate window into what you are doing and who you talk to — was never a network request. Nobody got to log it. There was no remote index to update with the fact that you are interested in forest gardening this week. The search did its work and left no trace anywhere but on your own machine.

Why the constraint makes it better, not worse

It is tempting to treat "runs locally" as a privacy tax — a worthy limitation you accept in exchange for losing some capability. I have come to see it the other way.

A search index that lives on a server can only ever see the slice of your life you were willing to upload to it. A search that runs on your device can draw on everything you have, because all of it is already there and none of it had to be risked by sending it away. The richest possible map of your digital life is one assembled in the one place where all of it can safely coexist — your own hardware. The connections that make a search feel like it understood you are exactly the connections you would never want to hand to a remote index in the first place.

So the constraint is not "good search, minus privacy." It is search that can afford to know more about you precisely because it never tells anyone. It understands what you meant, it follows how your memories connect, and it does all of it without the query ever leaving the room. That is what it means for search to never phone home.

Related: Memory That Only You Can Read · Memories That Remember the Path, Not Just the Point.

Written by AI agents from real project logs; owned and edited by Mujo.