It started with a file I could not find.

The file

A colleague sent me a file. A budget with a chart in it, some time in the spring. I searched Gmail by the words I remembered and got nothing: it had arrived on Slack. I searched Slack and got nothing: wrong month, wrong person. It had been in Finder for weeks, under a name I would never have typed.

Twenty minutes, three good search boxes, each blind to the others, none willing to take what I actually had: a half memory with an error in it. That evening I wrote the first version of Memoria. It ran the same vague query against every drawer at once and relaxed the detail that did not match.

Why local-first

The obvious way to build this is a cloud service that ingests your accounts. It is the wrong way, for two reasons that get stronger every year.

Hardware. A recent laptop runs a small language model and an embedding model comfortably. The compute that used to justify a server is under your hands. AI is moving home the way music and photos did, and search over your own life should make the trip first.

Sensitivity. An index of your email, files, chats and browsing is a profile, more sensitive than any single source. It should exist in one place, on your disk, and that should be provable. Local-first is not a feature. It is the constraint everything was designed under, and the verify page is where it is shown rather than claimed.

The arc

Memoria remembers for you. It finds the thing across every source, the way you remember it, wrong details included, and shows where it was and why it matched. Read-only, local, checkable. This is the product on the waitlist.

Replica works like you. A system that knows how you find things starts to know how you work: which sources you trust, how you name things, who you go to for what. Next is a local assistant that acts inside your own tools with your own habits. Still on your machine, still read-only until you decide otherwise.

Persona is yours to license. Your professional way of working becomes an asset with an owner, and the owner is you. Not a model trained on your output by someone else: a portable persona you hold, inspect, and license on your terms. Each step depends on the previous one being trustworthy. That is why the first product is so stubborn about proof.

On profiles

Most software that knows a lot about you knows it in order to sell it. The profile is the product; you are not its customer. We build the same depth of understanding with the opposite ownership: the profile is built on your machine, from your data, and stays yours. Nobody here can see it. Knowing you is not our business model. Finding your things is.

The founder

One technical founder, building in public. Every release is in the changelog, including what did not work. [PASTE: optional name and link]