Skip to main content

Memory Configuration

Configure the memory system to store and retrieve persistent information about user preferences, context, and important facts across conversations.

Complete Reference

For complete field documentation, see agent.yml → Memories.

Overview​

The Memory system allows your Dexto agent to remember information across conversations and sessions. Memories can be created by users, the system, or programmatically through the API.

Key features:

  • Persistent storage across sessions
  • Tagging and metadata support
  • Pinned memories for auto-loading
  • Flexible filtering and retrieval
  • Integration with system prompts

Memories are stored in your configured database backend and can be automatically included in the system prompt for context-aware interactions.

Memory Structure​

Each memory contains:

  • content - The actual memory text (1-10,000 characters)
  • tags - Optional categorization (max 10 tags, 1-50 chars each)
  • metadata - Source tracking, pinning, and custom fields
  • timestamps - Creation and last update times
# Example memory structure
content: "User prefers concise responses"
tags: ["preference", "communication"]
metadata:
  source: user
  pinned: true
  customField: "any value"

Enabling Memories​

Memory is configured at the top level of your agent.yml file:

memories:
  enabled: true
  priority: 40
  limit: 10

Configuration Options​

OptionTypeDefaultDescription
enabledbooleanfalseEnable memory inclusion in system prompt
prioritynumber40Position in system prompt (lower = earlier)
limitnumber-Maximum memories to include
includeTimestampsbooleanfalseShow last updated date
includeTagsbooleantrueInclude associated tags
pinnedOnlybooleanfalseOnly include pinned memories

Pinned Memories​

Pinned memories are automatically loaded into the system prompt when pinnedOnly is false, or exclusively loaded when pinnedOnly is true.

Configuring Pinned Memories Only​

memories:
  enabled: true
  priority: 40
  pinnedOnly: true    # Only load pinned memories
  limit: 5            # Keep system prompt compact

When to Pin Memories​

Pin these:

  • Critical user preferences (communication style, constraints)
  • Important project context (tech stack, standards)
  • User-specific requirements (accessibility needs, language)

Don't pin these:

  • Temporary context (current task details)
  • Historical information (past interactions)
  • Optional details (nice-to-have context)

Use Cases​

ScenarioMemory Strategy
Personal AssistantPin schedules, preferences, important dates
Customer SupportStore customer history, preferences, past issues
Development AssistantRemember tech stack, coding standards, project structure
Research AgentTrack research topics, sources, findings

Configuration Examples​

Basic Memory Integration​

llm:
  provider: anthropic
  model: claude-sonnet-4-5-20250929
  apiKey: $ANTHROPIC_API_KEY

systemPrompt: |
  You are a helpful AI assistant that remembers user preferences.

memories:
  enabled: true
  priority: 40
  limit: 15
  includeTimestamps: true
  includeTags: true

Hybrid Approach: Pinned + On-Demand​

memories:
  enabled: true
  pinnedOnly: true    # Only auto-load critical context
  limit: 5            # Keep system prompt compact

Then query additional memories programmatically when needed:

// In your application code
const memories = await agent.memory.list({
    tags: ["customer", "billing"],
    limit: 20
});

Example Output Format​

Memories appear in the system prompt as:

## User Memories
- User prefers concise responses [Tags: preference, communication]
- Project uses TypeScript with strict mode [Tags: technical, configuration]
- User's timezone is PST [Tags: personal] (Updated: 1/15/2025)

Storage Requirements​

Memories require persistent storage:

storage:
  database:
    type: sqlite  # Required for persistent memories

Memory data uses the key pattern: memory:item:{id}

Best Practices​

  1. Pin sparingly - Only pin critical information that should always be available
  2. Tag consistently - Develop a tagging strategy for easy filtering
  3. Keep content focused - Each memory should contain a single, clear piece of information
  4. Use source field - Track whether memories came from users or system
  5. Set reasonable limits - Use limit option to prevent system prompt bloat
  6. Regular cleanup - Review and remove outdated memories periodically
  7. Combine approaches - Use pinned for core context, query on-demand for specific needs

See Also​