Autopilot Memories

Telemetry tells you what is broken, but resolving complex incidents often relies on context that live data cannot provide. This vital context exists as operational or tribal knowledge—past failure patterns, effective workarounds, and unique infrastructure quirks—that is typically lost across fragmented daily interactions. Autopilot memory systematically extracts these operational lessons from your interactions with the Autopilot and transforms them into a persistent knowledge base accessible in the memory management portion of the New Relic UI.

Opt in

Autopilot memory requires an Advanced CCU SKU and access can be managed through Feature Control Manager. By default, Autopilot Memory is disabled. To opt in, an Organization Manager or Auth Domain Manager must explicitly enable the feature via the Feature Control Manager under Administration. This helps ensure your organization retains control over data capture and sharing.

How it works

If enabled, Autopilot memory continuously extracts knowledge from agent-user interactions to deepen Autopilot context and understanding to help improve response accuracy.

To provide the best answers, Autopilot relies on two types of memory—combining the immediate context of your current chat with historical knowledge saved from past sessions:

• Active thread context: Active chat sessions are temporarily retained for immediate context for quick follow-ups, allowing Autopilot to help maintain the continuity of your current investigation.
• Long-Term Memory: This memory is a persistent cross-session knowledge that allows Autopilot to have historical context around your preferences and your unique environment. Instead of explaining your setup to the agent in every chat session, Autopilot can automatically retrieve historical insights, user preferences, and past feedback in long-term memory. For example, if a previous investigation determined that a recurring database spike is expected behavior caused by a scheduled nightly backup, Autopilot can apply that operational context to future analyses. Long-Term Memory is categorized into three distinct scopes:
  • User Scope: Context and preferences saved exclusively for your individual interactions with the AI.
  • Account Scope: Shared operational knowledge accessible to all authorized users within a specific New Relic account.
  • Organization Scope: Global operational knowledge accessible to all authorized users across your entire New Relic organization.

You can view and manage these distinct scoped memories in the Memory Management UI.

Autopilot memory capture and retrieval

Data flows into and out of the long term memory through the following mechanisms:

• Automatic Memory: Once an Organization Manager enables memory via the Feature Control Manager, the system continuously processes your chat interactions behind the scenes. This automated pipeline helps Autopilot to retain incident context without requiring manual effort:
  • Fact and Relation Extraction: The system analyzes prompts and the responses to intelligently extract durable operational details, user preferences, and architectural context. These insights are automatically categorized and stored as structured facts and relations, ready for fast, semantic retrieval during future investigations.
• Manual Promotion (Save Response button in Chat Panel): You can also use the Save Response button in the chat panel to manually add what you may consider to be a valuable insight to the Autopilot memory. Your options for where to save this context depend on your RBAC permissions:
  • Organization Managers: Can choose to promote the insight to the Organization, Account, or User memory.
    • (Note: Apart from the ability to view and delete their own User-scoped memories, Organization Managers have full authority to view and manually delete memories at the Organization and Account scopes to keep the long-term memory accurate)
  • All Product Admins: Can choose to promote the insight to the Account or User memory.
    • (Note: Apart from the ability to view and delete their own User-scoped memories, All Product Admins have full authority to view and manually delete memories at the Account scope).
  • All other roles: Can save the response exclusively to their own private User memory.
• Feedback: Autopilot has the ability to adapt logic for your use over time based on the corrections, feedback, and context you provide. For example, when refining agent logic, providing highly specific feedback (e.g., "Update this NRQL filter to account for the auth-service memory leak") creates high-quality entries in Memories.
• Intelligent Retrieval: Rather than flooding prompts with unnecessary data, Autopilot dynamically selects the context. It uses vector similarity search to rank and retrieve only the most relevant historical lessons. It isolates memory to the active thread for immediate follow-ups, but can recall past sessions when you reference historical events.

Targeted memory deletion

While automated background jobs handle data retention, we recognize that your environment can change, sometimes rapidly. If a previously saved context becomes outdated (e.g., an architectural change makes an old workaround obsolete) or was captured incorrectly, authorized users can manually delete specific long term memory entries directly from the Memory Management UI. This helps to ensure the persistent long term memory base reflects your current operational reality.

Autopilot memory management UI

To use this feature, an Organization Manager or Auth Domain Manager must first enable it via the Feature Control Manager under Administration.

The centralized Memory Management dashboard gives you visibility and control over what Autopilot can use for context and understanding your environment. Once the feature is enabled, you can view, audit, and manage all memory entries here based on the RBAC permissions.

• Access: Access the dashboard from the left navigation menu under Knowledge > Memories.
• Scope Filtering: Easily toggle between Organization Level, Account Level, and User Level tabs to view scoped knowledge.
• Search Memories: Use the search bar for text search or apply filters to find specific contexts.
• Data Hierarchy: Memories are grouped by conversation threads. Expanding a conversation thread reveals granular details, including the facts and relations captured along with other metadata. Users can use this to view and manage outdated entries.

Autopilot Memory Role-Based Access Control (RBAC)

Access to the Memory Management UI is governed by your assigned role:

• Organization Managers: Can view and manage (create and delete) Organization, Account, and User Memory.
• All Product Admins: Can view Organization Memory, and manage (create and delete) both Account and User Memory.
• All other roles: Can view shared Organization and Account Memory, and manage (create and delete) their own private User Memory.

If you need to set up custom permissions for managing Autopilot memories, see Create custom roles for Autopilot Memories.

Data retention

All Long-Term Memory—whether scoped to a user, account, or the entire organization—is governed by what you have chosen for New Relic AI Data Retention policy in the Retention UI.

• Default Retention: The base retention period for Long-Term Memory is 90 days.
• Customization: You can modify this duration (up to a maximum of 365 days) by navigating to Administration > Data Retention, selecting the Organization-scoped tab, and adjusting the policy for the Long Term Memory source.

Next steps