One key per scope
The shared brain and every person have their own data key, stored wrapped and rotated by version. Pages, sources and facts are sealed under the key of the scope they belong to.
Security and privacy
Living Brain reads with the asker's own access. Grants, roles, agent keys, the audit log and encryption at rest per scope are live. Hidden-text screening, mail and bring-your-own key are planned, and the limits are stated plainly.
| Asked in | Shared brain | Private channel memory | Personal memory |
|---|---|---|---|
| Public channel | Yes | No | No |
| Private channel | Yes | That channel | No |
| Your DMs | Yes | Channels you're in | Yes |
Always limited to what the person asking can already see. It follows your chat's permissions (Slack channels, Discord roles), which are planned. In the web app and over MCP, reads follow the grants an owner or admin sets for each person, role, agent account, folder or project. How grants work.
The access rules above aren't only a check in code. Each scope has its own data key, so a leaked key exposes that scope and nothing else. Page bodies and fact values are sealed with XChaCha20-Poly1305 under the key of their scope.
The shared brain and every person have their own data key, stored wrapped and rotated by version. Pages, sources and facts are sealed under the key of the scope they belong to.
Search uses a blind index: keyed hashes of the words in each scope, so a query only touches the scopes you can see. Matching pages are decrypted in memory for that one request.
Planned: the CLI and the app keeping scopes under a key from your keychain or passkey, and searching locally without a network round trip. Today the service holds the keys.
Forget a channel, offboard a person or revoke a mailbox: delete that scope's key and its memory is unreadable everywhere, backups included.
The honest limit: the service holds the wrapping key and decrypts to serve and answer, so this is not end-to-end encryption, and the model reads the text to write the wiki. Slugs, subjects and attribute names are stored in plain text. Hosted plans protect your data at rest and scope by scope, but not from the running service itself. For "even you can't read it", self-host on the Community or Teams plan. Keyword hashes also reveal which words repeat within a scope, never across scopes.
Everything it reads, from Slack and Discord, mail, imports, agent logs and git, is screened for hidden text by PromptDecode, a sister Factory Zero venture, so an invisible instruction in one message can't reach every agent's context. Anything that decodes to an instruction is held for a person to review.
Connected mailboxes (Gmail, Outlook, IMAP) stay in the owner's personal memory, encrypted with their key, unless they share a specific label or folder. Every message passes Owlpost's spoofing and prompt-injection checks first, and mail can never trigger actions. Revoke a mailbox and its stored mail is erased.
livingbrain telemetry off turns it off.Today the marketing site and the hosted app run on Cloudflare. Everything else here is planned, and labelled so.
Factory Zero ventures this site uses or will use: Cratefield (framework, planned); Owlpost (email, planned); SupportGenius (support, planned); SupportGenius (bug reports, planned); promptdecode (security screening, planned); Keep Shipping (deploys, planned). The full list, with what is live and what is planned, is in the Factory Zero registry.