Persistent Memory

The Persistent Memory block stores and retrieves information across workflow sessions, enabling agents to remember context, learn from interactions, and maintain long-term knowledge.

Overview

Persistent Memory creates long-term information storage that survives beyond individual workflow executions. Unlike temporary variables that reset with each run, persistent memory maintains context across sessions, users, and time periods.

Information Storage: Save important context, decisions, and learnings to permanent storage

Context Retrieval: Access relevant historical information when processing new requests

Memory Organization: Structure information using tags, categories, and relationships

Memory Management: Update, merge, and clean up stored information over time

How It Works

graph LR
    A[New Information] --> B{Store or Retrieve?}
    B -->|Store| C[Memory Encoding]
    B -->|Retrieve| D[Memory Search]
    C --> E[Persistent Storage]
    D --> F[Context Matching]
    F --> G[Retrieved Memory]
    E --> H[Confirmation]

Configuration

Memory Store

The backend storage system for persistent data:

Vector Database: Semantic similarity search (Pinecone, Qdrant, Chroma)
Graph Database: Relationship-based storage (Neo4j, ArangoDB)
Key-Value Store: Simple structured storage (Redis, DynamoDB)

Memory Scope

Defines what information gets stored and retrieved:

User-Specific: Individual user contexts and preferences
Global: Shared knowledge across all users
Session-Specific: Conversation threads and temporary context

Retention Policy

Rules for memory lifecycle management:

Time-Based: Delete memories older than specified duration
Usage-Based: Remove rarely accessed information
Capacity-Based: Maintain maximum storage limits

Relevance Threshold

Minimum similarity score for memory retrieval, preventing irrelevant context pollution.

Use Cases

Customer Service: Remember previous interactions, preferences, and issue history
Personal Assistants: Maintain user preferences, schedules, and personal context
Knowledge Management: Build organizational knowledge bases that grow over time

Example Workflow

[User Message] → [Persistent Memory] → [Agent] → [Update Memory] → [Response]

A customer service scenario:

First Interaction: Customer asks about return policy for electronics

Store: Customer interest in electronics, policy questions
Response: Detailed return policy information

Second Interaction (weeks later): Customer asks about warranty on tablets

Retrieve: Previous electronics interest, policy questions
Context: "I see you were asking about electronics returns before..."
Enhanced Response: Warranty info + proactive return policy reminder

Memory Evolution: System learns customer is electronics-focused and policy-conscious

Best Practice: Design clear memory categories and use semantic embeddings for flexible retrieval. Regularly audit and clean stored memories to maintain relevance.

When to Use This vs Other Blocks

Block	When to Use
Persistent Memory	Long-term context that needs to survive across sessions
Episodic Memory	Sequential conversation history and recent interactions
Variables	Temporary data within a single workflow execution

Persistent Memory

On this page