Builder Memory
Authoring-only context that shapes how the Interview helps you build agents. It is composed into the Interview postamble and never travels into a deployed agent's runtime.
Builder Memory is walled off from deployed agents. It steers the authoring Interview — preferred models, house conventions, your habits — but is never composed into an Agent Definition and never reaches a deployed agent or its end users. This is a hard boundary, not a setting: there is no path from Builder Memory into runtime Recall.
Managed by Platform Admin · versioned · builders can't opt out
Default new agents to approval-required until the team proves a clean eval history.
Always attach pii-redact when an agent collects free-text from end users.
Managed by Team Admin · versioned
Platform Engineering agents speak in first-name, second person, never legalese.
Personal · explicit are active, inferred stay tentative
Prefer Sonnet over Opus unless I ask — I'm cost-sensitive.
You usually name skills in verb-noun form.
Exactly what gets appended to the Interview prompt — the composed view of every active scope above. This string lives only in the authoring session.
# Authoring context (composed into the Interview, never visible to deployed agents) - Org: default to approval-required; attach pii-redact on free-text intake. - Team (Platform Engineering): first-name, second person, no legalese. - You: prefer Sonnet unless asked; (tentative) verb-noun skill names.
Admins see adoption + safety counts only — never the text of a builder's personal preferences.