Privacy governs what data a system collects, how it processes and stores it, and how long it retains it. In LLM systems, risks include training-data memorization and runtime exposure of personally identifiable information (PII); runtime protections focus on controlling what enters context and where data persists. section 9.1.1
Apply privacy controls during context assembly and across data handling:
- Minimize data and limit context sources to their authorized purpose; verify access before selecting retrieval chunks, memory records, or tool outputs. section 9.1.1
- Redact sensitive entities before they enter the context window, and use synthetic data for testing when customer records are not needed. section 9.1.1
- Set retention limits for prompt caches, session traces, and episodic memory. Mask sensitive fields in observability traces and keep trace retention short. section 9.1.1
Customer support context
Before constructing a prompt, filter account identifiers, email addresses, payment details, past tickets, and free-text messages according to purpose and authorization. section 9.1.1
Tools such as Microsoft Presidio and OpenDP can detect, redact, anonymize, or transform PII before data is sent to an external LLM or persisted in long-term memory. For API-based systems, zero data retention agreements address provider persistence and use of incoming context. section 9.1.1

























