Conversations
Conversations are kept for 30 days and then deleted. They are never used to train models. Each person’s conversations are separate; nobody else on the team can read yours through the assistant.
Before a message is stored, secrets in it are replaced with a placeholder. If a message contains a private key, the whole message is dropped and the assistant says so. Do not paste secrets into the assistant.
What the assistant never sees
The assistant never sees a credential. Calls to the Overlane API go through a credentialed proxy that attaches your permissions on the server, so no token appears in a prompt, a log, a stored message or the page. The proxy refuses any path under another team, and any change that does not come from an approved write.
Text from issues, pull requests, docs pages and other retrieved content is treated as data. It is quoted and labelled as quoted data and is never read as an instruction.
The audit log
Every tool call is recorded, including refusals: who asked (user and role), which tool, a digest of the arguments, and the result. The log holds a digest, not the arguments and not the text of any message. Approvals record who proposed and who decided. Admins and owners can read the team’s log.
The docs panel
The panel on this site answers in your browser from the docs search index. What you type there is not stored and is not sent anywhere.
Models
Docs answers and navigation use search and rules, not a model. A team may configure a planner model for requests that no rule understands; it sees only the words you typed and the list of tools, never retrieved text or a credential, and what it proposes still needs your approval.