The week you just had,
kept.
Here is a working week. A scope decision in ChatGPT. A fix worked out in Cursor. A note in Obsidian, a calendar move, something worth keeping from email. By Friday it is scattered, and by next month most of it is gone. This page is what changes: the features below are the ordinary moments of that week, and what each one turns into when the memory is shared.
Claude does not know what you told Cursor.
The briefing you gave ChatGPT on Monday is there when Cursor needs it on Thursday, source attached. A decision you worked out in March is still answering questions in July. Nothing gets re-explained.
One MCP address; every client reads and writes the same brain.
See how clients connectSee the record of this site’s own build
You know you saved it. You can’t remember how you put it.
Ask for what you meant rather than what you typed, and the note turns up. The version of a fact you need most, usually the latest, ranks first, and old versions stop drowning it out.
Semantic search over stored meanings, with time-decay reranking.
Read how recall worksYou asked, and got back a pile of links.
Ask a real question and the answer is built from what you actually stored, with the memories that support it listed underneath. Open any of them and check it against the source.
Recall and synthesis run on Workers AI inside your account.
See the dashboardThe meeting moved. Now your notes disagree.
The brain flags the conflict and settles on the newer belief, keeping the older one for audit. Two half-notes about the same thing become one note. Save the same article twice and no copy piles up.
Captures are compared by meaning before storing: 85 to 95 per cent similarity goes to a merge decision, 95 and above is blocked.
Read how conflicts are settledThe thought arrived in the wrong app.
Save it where it happened: the terminal, the share sheet, the bookmarklet, the Obsidian plugin. Calendar and inbox sync read-only, so the week’s movements land by themselves.
Capture surfaces from the terminal to the share sheet, all writing to one store.
See every capture surface and sync sourceWhere did that come from?
Each memory shows where it was saved from, and a hint appears when a fact may have gone stale. High-importance entries are protected from being quietly overwritten.
A source field on every capture, and a staleness pass over the store.
Read the source and staleness fieldsThe same theme keeps turning up.
Threads that recur across what you have stored surface as patterns, and a weekly note gathers them. You stop compiling your own themes by hand on a Sunday evening.
Pattern derivation runs after every search; insights ship weekly.
See weekly insights in the dashboardYou have abandoned tools like this before. Setup was the reason.
The desktop app asks for a Cloudflare login and a password, then provisions the database, the vector index and the Worker, and wires up Claude Code and Cursor. About two minutes. No terminal.
Signed builds for Mac and Windows; the app can update the deployed Worker in place.
Read the setup guide
Shared on purpose,
private by default.
Team Edition runs on the same Worker as the personal brain, the same deployment and no extra charge, and nothing is shared until someone shares it. Each person has two layers: a personal layer nobody else can see, search or reach through any client, and the shared layer everyone on the team reads. Sharing moves the memory itself, one entry with one history, not a copy, and it can be moved back; once shared, only its author or an admin can change, un-share or delete it.
Every member signs in with their own one-time token. It is shown once, stored only as a hash, and nobody can retrieve it afterwards, the admin included; resetting it disconnects that one person and leaves everyone else working. Identity follows the token, so each person’s clients search their personal layer plus the shared layer, and members never need a Cloudflare account of their own.
Admins manage members and roles, reset tokens, set capture defaults, turn team insights on or off, and review an activity feed of every add, share and removal, exportable as CSV. What they cannot do is read anyone’s personal layer, the owner’s included. For offboarding they see a count of a person’s private memories, never the content.
Suspension blocks access and deletes nothing; it is reversible. Removal is permanent: that person’s personal memories go, while anything they shared stays, still credited to their name. You cannot remove yourself, or the last admin.
Limits. One shared team per brain in v3.0.0, with no enforced size limit; the practical ceiling is your Cloudflare plan’s budgets. A personal brain becomes a team brain when a second person joins: existing memories are filed into your personal layer and nothing is shared automatically. Going back to solo is possible but has a cost: removing everyone deletes each person’s personal memories, while the shared layer stays.
Start with your own week.
Two minutes, your Cloudflare account, free and MIT-licensed forever. There is no paid tier, and no paid team tier either.