Agent Memory

Memory allows agents to remember previous interactions and maintain context. Without memory, each agent interaction starts fresh. With memory, agents learn and adapt over time.

Memory Types

Conversation Memory (Short-term)

Remembers recent messages and context in the current conversation.

Configuration:

• Lookback window - How many previous messages to remember
• Token limit - Maximum context size
• Reset behavior - When memory clears

Best for: Customer support chats, multi-turn conversations

Task Memory

Tracks progress and state during multi-step workflows.

Best for: Complex decision-making, approval workflows, multi-step processes

User/Context Memory (Long-term)

Remembers user preferences and patterns across sessions.

Configuration:

• Retention period - How long to keep information
• What to remember - User preferences, patterns, history

Best for: Personalization, adaptive behavior

Memory Scope

Execution-Level

Agent only remembers its own interactions within a single workflow run.

• Resets between workflow executions
• Lightweight and fast

Workflow-Level

Agent has access to entire workflow context and decisions from other nodes.

• Agent sees what previous nodes output
• Useful for complex workflows needing full context

Configuring Memory

When setting up an agent node:

1. Open agent configuration
2. Go to Memory section
3. Select memory type(s) you want: Conversation, Task, User memory
4. Configure parameters (lookback, token limits, retention)
5. Choose scope: Execution-level or Workflow-level
6. Save

Best Practices

• Match to use case - Use conversation memory for chats, long-term for personalization
• Start minimal - Add memory only when needed
• Monitor consumption - Large context windows slow down processing
• Test thoroughly - Use Playground to validate memory behavior
• Balance context and performance - More memory = better context but slower

Testing Memory

Use the Agent Playground to validate:

• Does agent remember previous context?
• Does memory improve agent behavior?
• Is performance acceptable?
• Does memory scope work as expected?

See Agent Playground documentation for testing guidance.