Long-term memory categories
Summarize
Summary of Long-term memory categories
Long-term memory (LTM) categories in Now Assist AI define the types of semantic information an AI agent can learn and retain about users over time. These categories enable personalized agent responses by accumulating user-specific context across interactions. LTM categories are organized in the AI Agent Memories table and can be customized and mapped to specific agents to control what information is stored and retrieved.
Show less
Key Features
- Semantic memory organization: User-specific information is categorized (e.g., software preferences, work context, communication style) and stored as semantic memories scoped by user and category.
- Category mapping: LTM categories are defined globally and mapped to agents, allowing agents to learn and recall only relevant categories to personalize interactions without requiring repeated user input.
- Default categories included:
- Software and tools: Tracks user applications, software configurations, and related preferences.
- Work context: Captures user role, department, location, and workplace preferences.
- User preferences: Stores communication and interaction style preferences like language and notification settings.
- Custom categories: Organizations can create additional categories tailored to their specific use cases.
- Memory extraction process: An LLM prompt evaluates interactions to extract user facts matching mapped categories, which are then stored as JSON objects in the AI Agent Memories table for retrieval during future interactions.
Considerations for ServiceNow Customers
- Use clear, specific category descriptions to improve the accuracy of semantic memory extraction and reduce false positives or missed data.
- Limit the number of categories mapped to an agent to avoid increased processing time during memory extraction.
- Verify extracted semantic memories by filtering the AI Agent Memories table for type "Semantic" and by relevant agent or user to ensure the AI agent is effectively learning user context.
Benefits
By leveraging LTM categories, ServiceNow customers can enhance their AI agents to provide more relevant, context-aware responses, improving user experience and reducing repetitive queries. This capability supports personalized assistance aligned with user roles, preferences, and work environments, driving more efficient and effective automation.
Long-term memory (LTM) categories define the types of semantic information that a Now Assist AI agent can learn and retain about users over time. You can add new categories and map them to specific agents to personalize agent responses based on accumulated user context.
Semantic memory in the AI agent Memories table (sn_aia_memory_list) is organized by LTM categories. Each category represents a distinct type of user-specific information, such as software preferences, workplace
context, or communication style. By mapping categories to an agent, you control what the agent learns and retains across interactions.
How LTM categories work
When an agent executes, the platform evaluates the interaction for user-specific facts that match the agent's configured LTM categories. Matched facts are stored as semantic memory records in the AI Agent Memories [sn_aia_memory] table, scoped to the user and category. On subsequent interactions, the agent retrieves relevant semantic memories and uses them to personalize its responses without requiring the user to repeat context.
LTM categories are defined globally and can be mapped to one or more agents. An agent only learns and retrieves memories for the categories explicitly mapped to it.
Default LTM categories
The platform includes the following default LTM categories:
- Software and tools
- Captures information about applications, tools, and software configurations relevant to the user, such as operating system version or approved applications.
- Work context
- Captures facts about the user's role, department, location, and workplace preferences, such as remote work setup or team structure.
- User preferences
- Captures communication preferences and interaction style, such as preferred language, response format, or notification settings.
You can extend this list by creating custom categories suited to your organization's use cases.
How categories affect memory extraction
During memory extraction, the platform runs an LLM prompt that evaluates the agent interaction against the descriptions of each mapped LTM category. If the interaction contains information that matches a category, a semantic memory record is created or updated in the AI Agent Memories table with the following fields:
- Category
- The LTM category that the memory is associated with.
- User
- The user whose interaction generated the memory.
- Memory
- The extracted user-specific fact, stored as a JSON object.
- Type
- Set to Semantic for category-based memories.
Semantic memories are retrieved at runtime using retrieval-augmented generation (RAG) and injected into the agent prompt to personalize the response for the current user.
Considerations
- Category descriptions directly influence LLM extraction quality. Use specific, unambiguous language to reduce false positives or missed extractions.
- Mapping too many categories to a single agent can increase extraction processing time. Map only the categories relevant to the agent's use case.
- To verify that semantic memories are being extracted, open the AI Agent Memories table and filter by Type = Semantic and the relevant agent or user.