Memory-Aware Chatbot with Ollama and Mem0

A sophisticated chatbot that combines Ollama's local LLM capabilities with a custom memory system (mem0) for persistent, context-aware conversations. Features advanced memory management with extraction, update, and retrieval phases powered by LangChain and vector similarity search.

Features

• 🧠 Persistent Memory: Remembers information from previous conversations with intelligent storage
• 🔍 Semantic Search: Finds relevant memories using FAISS vector similarity search
• 💬 Natural Conversations: Context-aware responses using retrieved memories and conversation history
• 📚 Advanced Memory Management: Intelligent ADD, UPDATE, DELETE, and NOOP operations
• 🤖 Local LLM: Uses Ollama with LangChain for privacy-focused AI interactions
• ⚡ Multi-Phase Architecture: Extraction → Update → Retrieval pipeline for memory management
• 🔄 Automatic Memory Extraction: LLM-powered extraction of important facts from conversations
• 📊 Memory Analytics: Search, view, and manage stored memories with similarity scoring

Prerequisites

1. Python 3.8+
2. Ollama: Install from https://ollama.ai
3. LangChain: For LLM integration (included in requirements)
4. FAISS: For vector similarity search (included in requirements)

Quick Start

1. Setup Ollama

# Install and start Ollama (if not already done)
ollama serve

# In another terminal, pull a model
ollama pull qwen2.5:3b-instruct

# run the model in other terminal
ollama run qwen2.5:3b-instruct

2. Install Dependencies

# Install Python dependencies
pip install -r requirements.txt

Or run the setup script for automated checking:

python setup.py

3. Run the Chatbot

python chat.py

Usage

Basic Chat

Just type naturally and the chatbot will respond while learning about you:

👤 You: Hi, I'm working on a React project
🤖 Assistant: Hello! That's great that you're working on a React project...

Special Commands

• memories - View recently stored memories with details
• search: <query> - Search memories using semantic similarity
• help - Show all available commands
• quit / exit / bye - End the conversation gracefully

Advanced Features

The chatbot automatically:

• Extracts memories: Identifies important facts from conversations
• Updates existing memories: Merges or updates related information
• Removes redundant data: Avoids storing duplicate information
• Maintains context: Uses recent conversation history for better responses

Example Session

👤 You: I love playing Valorant, especially as Sage
🤖 Assistant: That's awesome! Sage is a great agent...

👤 You: memories
📚 Recent Memories:
- User enjoys playing Valorant and prefers playing as Sage agent

👤 You: search: gaming
🔍 Searching memories for: 'gaming'
1. Score: 0.892
   Content: User enjoys playing Valorant and prefers playing as Sage agent

Architecture

Memory Management Pipeline

The system follows a sophisticated 3-phase architecture:

1. Extraction Phase (extraction.py): Analyzes conversations and extracts important facts
2. Update Phase (update.py): Determines appropriate memory operations (ADD/UPDATE/DELETE/NOOP)
3. Retrieval Phase: Searches and retrieves relevant memories for context

Components

1. Database (database.py): Handles memory storage, vector operations, and FAISS indexing
2. Extraction (extraction.py): LLM-powered extraction of important information from conversations
3. Update (update.py): Intelligent memory management with conflict resolution
4. OllamaLLM (ollama_wrapper.py): LangChain-based wrapper for Ollama API integration
5. MemoryAwareChatbot (chat.py): Main orchestrator coordinating all components
6. Prompts (prompts.py): Centralized prompt templates for consistency

Memory System

• Storage: JSON files for memories and conversation history with structured metadata
• Vector DB: FAISS for high-performance semantic similarity search
• Embeddings: Uses Ollama's embedding models (nomic-embed-text) for vector representations
• Extraction: LLM-powered fact extraction with context-aware prompting
• Updates: Intelligent memory operations to prevent redundancy and maintain accuracy
• Conflict Resolution: Automatic handling of contradictory or duplicate information

Architecture Diagram

Database structure

Extraction Phase

File Structure

mem0/
├── chat.py              # Main chatbot application with interactive loop
├── ollama_wrapper.py    # LangChain-based Ollama API wrapper
├── database.py          # Memory storage and FAISS vector operations
├── extraction.py        # Memory extraction logic with context assembly
├── update.py            # Memory update phase with intelligent operations
├── prompts.py           # Centralized prompt templates
├── memories.json        # Stored memories with metadata
├── message.json         # Conversation history and context
├── summary.txt          # Conversation summary for context
├── memory_embeddings.json # Memory embeddings cache
├── memory_index.faiss   # FAISS vector index for similarity search
├── requirements.txt     # Python dependencies including LangChain
├── setup.py            # Setup script with dependency checking
├── images/              # Directory for architecture and demo images
│   ├── database.png         # System database diagram
└── README.md           # This file

Configuration

Changing Models

Edit the model name in chat.py:

chatbot = MemoryAwareChatbot(model_name="mistral")  # or any other Ollama model

Adjusting Memory Settings

In MemoryAwareChatbot.__init__():

• Modify similarity thresholds for memory retrieval
• Change context window sizes for conversation history
• Adjust memory extraction sensitivity and filtering
• Configure vector database parameters (dimensions, similarity metrics)

Memory Operations

The system supports four types of memory operations:

• ADD: Store completely new information
• UPDATE: Enhance existing memories with additional details
• DELETE: Remove outdated or incorrect information
• NOOP: No operation needed (information already exists or irrelevant)