Long-term memory categories
Summarize
Summary of Long-term memory categories
Long-term memory (LTM) categories in Now Assist AI enable agents to learn and retain specific types of semantic information about users over time. This personalization allows agents to provide tailored responses by leveraging accumulated user context without requiring repeated input from users. LTM categories organize user-specific information into distinct groups, which can be mapped to individual AI agents to control what they learn and recall across interactions.
Show less
Key Features
- Semantic Memory Organization: User facts are stored in the AI Agent Memories table, categorized by LTM types such as software preferences or work context.
- Configurable Category Mapping: Categories are defined globally and can be selectively assigned to agents, ensuring agents only learn and retrieve relevant memory types.
- Default Categories: The platform includes predefined categories like Software and Tools, Work Context, and User Preferences, which capture application usage, workplace roles, and communication styles respectively.
- Custom Categories: Organizations can extend default categories to suit specific use cases, enhancing personalization capabilities.
- Memory Extraction Process: During interactions, the platform uses LLM prompts to extract relevant user facts matching the mapped categories, storing them as semantic memory records with details on user, category, and memory content.
- Retrieval-Augmented Generation (RAG): At runtime, agents retrieve these semantic memories to personalize responses effectively.
Considerations for Implementation
- Use clear and specific language in category descriptions to ensure accurate memory extraction and minimize errors.
- Limit the number of categories mapped to each agent to optimize processing time and performance.
- To verify memory extraction, inspect the AI Agent Memories table filtered by semantic type and relevant user or agent.
Benefits for ServiceNow Customers
By leveraging LTM categories, ServiceNow customers can enhance their AI agents to remember and use user-specific details effectively, resulting in more personalized and efficient interactions. This capability reduces redundant user input, accelerates issue resolution, and improves overall user satisfaction with AI-assisted services.
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.