Search that covers both halves
One search box over your notes and your library — the documents, sheets, decks and PDFs, not just the pages you wrote. Instant, on your machine, and it works with the network off.
Know · the first of four
A workspace that can
find its own past.
Creating is the part everyone demonstrates. The part that decides whether a workspace is worth living in is whether you can get anything back out of it — six weeks later, when you have forgotten which file it was in.
Four things that make work findable
Links that hold
Write [[like this]] to connect one page to another, and the page you linked to shows you what points at it. Connections you make by hand survive being forgotten — you can walk them from either end.
A to-do list that agrees with your notes
Tick a box in a page and it is the same commitment everywhere it appears. Tasks are tracked by identity rather than by their position in the document, so reordering a page does not scramble what is done.
Asking in your own words
Semantic search finds the thing you can describe but cannot name. It needs an account and a network connection — the meaning-matching runs in the cloud today, not on your device.
Where the line is
“Local-first” is a claim worth being precise about, because most products that make it mean something weaker.
| Where it runs | Works offline | |
|---|---|---|
| Writing and editing | Your machine | Yes |
| Full-text search | Your machine | Yes |
| Wiki-links and backlinks | Your machine | Yes |
| Notes at rest | Encrypted with your OS keychain | Yes |
| Semantic search | Cloud — needs an account | No |
| Asking the agent a question | Cloud model, on the passages it needs | No |
| Transcribing a recording | Managed service — the audio is uploaded | No |
The last three rows are the honest ones. Ordinary work — writing, finding, linking, editing — happens on your machine and keeps working with the network off. The parts that need a model need the network, and when they run you can see exactly what crosses. We would rather print the table than say “private” and let you assume the rest.
What the agent already remembers
LumenAgent keeps durable memory across sessions — preferences, context about your work, things you have told it once and should not have to repeat. Credentials are screened out before anything is stored.
And what it does not. That memory has no source links and no screen to inspect it on, so it remembers what you tell it — not the decision you made in a document three weeks ago, with a citation. That is the difference between a helpful assistant and a workspace that understands your work, and it is the gap the next section is about.
What is being built next
This page is called Know because it is one of the four things the workspace does — not because it is finished. Here is the honest state of the rest, so you can decide whether to wait.
One item has left this list. Your day, in one place sat here until it was built. Today is the day surface and Calendar is the fortnight over Google and Outlook — with per-calendar and per-event exclusion from context, which is the part neither provider can give you. Everything below is still ahead.
- NextSearch that understands relationships, on your device
Meaning-matching computed locally instead of in the cloud, and the connections between your notes, files and tasks treated as something the search can follow. This is what makes the local-first claim cover semantic search too.
- ThenMeetings that produce work
A meeting as something the workspace understands, with decisions and action items that carry the point in the transcript they came from. Today a recording and its transcript attach to a note.
- ThenMemory & Context you can govern
A screen showing what is remembered, where it is stored, what AI is allowed to touch — and a Forget that removes it rather than hiding it. Plus a preview of exactly what a cloud request is about to send.
No dates, deliberately. We would rather tell you the order than a quarter we would have to quietly change.
Questions
Is my search index on my machine or in the cloud?
Full-text search is local and instant. The notes index is held in memory rather than written to disk, so encrypted note content never lands on disk in a searchable form. Semantic search — the kind where you describe a thing rather than name it — runs in the cloud and needs an account.
Are my notes encrypted?
Yes, at rest, using your operating system’s own keychain. That covers the note content on disk. It is not end-to-end encrypted sync — when you share a document, the keys are managed on the server so the recipient’s browser can render it.
Does the agent read my documents without being asked?
No. Searching the library is a tool it uses when your request needs it, and anything that leaves your machine passes an approval gate first. Automations watch for their trigger without spending AI credits and without reading anything until they fire.
Can I get my work out?
The files are on your disk in the first place. Every editor exports the formats other people expect, and notes export as Markdown or HTML.
Is any of this used to train a model?
No. No ads, no training on your content. When an AI step runs you can see what was sent, itemised.
The other three
Create — ten real editorsConnect — 108 apps, on your termsAct — hand over the outcome