Agentic Memory System

Inspired by Mem0: Building Production-Ready AI Agents with Scalable Long-Term Memory and Mem0 repo

This educational repository demonstrates an agentic memory system implementation based on the research from Mem0: Building Production-Ready AI Agents with Scalable Long-Term Memory. The system provides scalable, long-term memory capabilities for AI agents with intelligent fact extraction, semantic search, and memory consolidation.

Core Architecture

The agentic memory system follows the Mem0 architecture with the following key components:

Memory Storage Layer

• Vector Database: Qdrant for semantic similarity search
• Embedding Model: OpenAI text-embedding-3-small for vector representations
• Metadata Indexing: filtering by user
• CRUD Operations: Complete create, read, update, delete capabilities

Memory Processing Layer

• Fact Extraction: AI-powered extraction of factual information from conversations
• Memory Consolidation: Merging and updating of existing memories
• Semantic Search: Context-aware retrieval of relevant memories

Memory Management Layer

• CRUD Operations: Full create, read, update, delete capabilities
• Memory Actions: ADD, UPDATE, DELETE, UNCHANGED operations

Memory Algorithm

1. Fact Extraction

• Extracts factual information from natural language input
• Identifies personal preferences, plans, goals, and key information

2. Memory Consolidation

• Compares new facts with existing memories
• Determines appropriate actions (ADD/UPDATE/DELETE/UNCHANGED)
• Preserves important historical information
• Updates outdated information intelligently

3. Semantic Search

• Vector-based similarity search using embeddings
• Configurable similarity thresholds

Memory Operations

Adding Memories

await memory.add("I'm a software engineer at Akieni and I prefer TypeScript", {
  userId: "user123",
});

Searching Memories

const results = await memory.search("What programming languages do I know?", {
  userId: "user123",
});

Updating Memories

await memory.update("memory_id", "Updated memory content");

Deleting Memories

await memory.delete("memory_id");

System Components

Memory Class

The main interface for memory operations:

• add(): Extract facts and consolidate with existing memories
• search(): Semantic search with natural language queries
• get(): Retrieve specific memory by ID
• update(): Update specific memory content
• delete(): Remove specific memories

MemoryStorage Class

Complete storage implementation with:

• Vector Database Management: Qdrant integration with proper indexing
• Embedding Generation: OpenAI embedding model integration
• CRUD Operations: Full create, read, update, delete capabilities

Prompt Templates

AI-powered processing:

• extractFactsPrompt: Extracts factual information from input
• updateMemoryPrompt: Determines memory consolidation actions

🔧 Configuration

Environment Variables

QDRANT_URL=your_qdrant_url
QDRANT_API_KEY=your_qdrant_api_key
OPENAI_API_KEY=your_openai_api_key
MEMORY_COLLECTION_NAME=agentic_memory

Vector Database Setup

const storage = createMemoryStorage({
  url: process.env.QDRANT_URL!,
  apiKey: process.env.QDRANT_API_KEY,
  collectionName: "memory",
});
await storage.initialize();

Testing & Examples

The repository includes comprehensive examples:

• Basic Operations: Core memory operations
• Memory Consolidation: Memory merging

Run the examples:

npm run dev

References

• Mem0: Building Production-Ready AI Agents with Scalable Long-Term Memory
• Qdrant Vector Database
• OpenAI Embeddings

Credit

• Mem0: Building Production-Ready AI Agents with Scalable Long-Term Memory
• Mem0 Github Repo