
Overview
Memory in Relevance AI helps you extract and organize key information from conversations. These structured data points act as smart labels that make your conversations:- Searchable - Find exactly what you need
- Filterable - Sort conversations by specific criteria
- Consistent - Ensure the same data is used throughout conversations
- Intuitive - Easily recall and update memorized data when needed
Features
Add Short-term Memory
Adding Short-term Memory allows you to enrich and recall your data with additional context:- Navigate to the Build tab located at the top of your screen, adjacent to Tasks
- Click on “Memory” in the left sidebar navigation menu
- Choose your extraction method (Agent-decided or Rule-based)
- Define the appropriate data format and configure your settings
- Click the “Save” button to apply your changes
Consider what information would be most valuable for future analysis when deciding which metadata to add. Focus on data points that will help you segment, filter, or understand patterns in your agent interactions.
Delete Memory
To remove memory that’s no longer needed:- Find the memory field you want to remove in the Memory page
- Click the delete (trash) icon next to the field
Deleting memory is permanent and will remove this information from all associated conversations or tasks. Make sure you no longer need this data before proceeding.
Edit Memory Field
To modify an existing memory field:- Find the memory field you want to edit in the Memory page
- Click the edit (pencil) icon next to the field
- Update the field name, data format, or extraction method as needed
- Save your changes
Memory Extraction Methods
Relevance AI offers two approaches for extracting memory metadata:Let Agent Decide
With this option, you provide instructions to your agent about how and when to extract fields from memory:- How it works: You define guidelines for the agent to follow when determining memory values
- Best for: Complex scenarios where context-aware judgment is needed
- Example: “Extract customer sentiment as ‘positive’, ‘neutral’, or ‘negative’ based on the overall tone of the conversation”
Rule Based
This approach uses predefined conditions to automatically extract memory metadata:- How it works: You set specific triggers or patterns that, when detected, will populate memory fields
- Best for: Consistent, predictable data points that follow clear patterns
- Example: If a conversation contains the phrase “interested in pricing,” set lead_status = “pricing inquiry”
Data Formats for Short-term Memory
Relevance AI supports various data formats to accommodate different types of information:- Text
- Free-form text entries
- Useful for descriptions, notes, or other unstructured information
- Example: “Customer mentioned competitor Product X”
- Number
- Numerical values for quantitative data
- Supports integers and decimal values
- Example: Customer satisfaction score (1-10)
- True/False (Boolean)
- Binary values for yes/no or true/false conditions
- Example: “is_existing_customer”: true
- Single Option
- Predefined list where only one value can be selected
- Example: Lead status: [“New”, “Qualified”, “Opportunity”, “Customer”]
- Multiple Option
- Predefined list where multiple values can be selected
- Example: Products of interest: [“Product A”, “Product B”, “Product C”]
Choose the appropriate data format based on how you plan to use the metadata later. For example, if you want to calculate averages or totals, use the number format. If you need to filter by specific categories, single or multiple option formats work best.
Best Practices
- Be consistent: Use standardized naming conventions and values for your memory fields
- Start small: Begin with a few key memory fields and expand as needed
- Review regularly: Periodically assess which memory fields are providing value
- Document your schema: Keep track of what each memory field represents and how it should be used