Security and data
Your memory holds your decisions, your code’s history and sometimes your customers’ names. This page says where it is kept, what ever leaves it, and what stops an agent or a module from doing more than you allowed.
Where it is kept
Section titled “Where it is kept”- The cloud keeps it in an EU data centre. Your data shows the region, the last backup, and what is kept per workspace.
- Your own daemon keeps it on your machine, and nothing leaves unless you connect it: Where memory lives.
Deleting a workspace or the organization removes it from the live data at once and from the backups after 30 days.
What is masked before it is stored
Section titled “What is masked before it is stored”Keys, tokens, passwords, emails, phone numbers and public IP addresses are replaced with [REDACTED:<kind>] before a record or a conversation is stored, in every workspace of the cloud. On a daemon it is on by default, and you can add your own patterns: Redaction.
What leaves it
Section titled “What leaves it”Nothing, unless you turn it on:
| What | When | Where it goes |
|---|---|---|
| a search’s query and its best candidates | an accurate search, in a workspace that turned Voyage on | Voyage AI |
| what an agent sends to its model | an agent’s turn | the model provider whose key you entered |
| a call to an outside API | a call over a connection | that service, with your credentials |
| a mail | a send through Mail | your own mail provider |
Voyage is off in a new workspace. The models, the APIs and the mail are always your own accounts, never ours.
Keys and secrets
Section titled “Keys and secrets”- API keys are shown once, when made. Each has a scope (the whole organization, or chosen workspaces), read and write or read only, and an expiry; Revoke stops one at once: Connections and keys.
- Secrets in settings (a provider’s key, a mail password, an API token) are write-only: once saved, nobody sees them again, in the app or through the API. They are sealed for the organization and the one module that uses them.
- A connection’s credentials go only to its own host. An agent names the connection and never holds the token.
- Two-factor authentication protects your account, and an organization can require it: Profile.
What an agent may do
Section titled “What an agent may do”- An agent calls only the tools of the modules it may use. Tools only a person may call are withheld until a person grants them, one by one.
- A decision with a risk (code, money, data, a running system) waits for a person, unless a team’s trust allows that class, within limits a person set. No agent can raise its own trust or approve its own decision.
- Every decision, its approval and its note are kept, on the board and in memory.
What a module may do
Section titled “What a module may do”- Each module gets only what the organization grants it: to write memory, to reach the network, to read files, to start processes. Every call is checked against it.
- In the cloud every module runs in a sandbox of its own, without the network unless its grant gives it, and never reaching the cloud’s own network.
- A custom module has no network at all, no processes and no files outside its scratch folder, and runs only once a person approved exactly that code.
- A workflow calls only the modules a person approved for it.
Who looked at it
Section titled “Who looked at it”Every change to the organization’s modules is in the audit log on Team. Your data shows whether our support accessed your data this year, and links the full audit log and the data processing agreement.
Take it with you
Section titled “Take it with you”Export everything gives one file in the same format a daemon reads: your records, the graph, the workspaces, the cards and the vectors. Import it into a daemon on your own machine and it works there as it did here. Nothing locks your memory in.