LevTech GreatMinds #10 Talk & Demos

This repository demonstrates how to build real agentic AI systems using OpenAI Agents library, FalkorDB, and Graphiti—moving beyond simple prompt-response patterns to create agents with memory, knowledge, and reasoning capabilities.

These demos explore:

• Memory Graphs with Graphiti + FalkorDB - Episodic memory for conversational agents
• Knowledge Graphs with FalkorDB - Structured domain knowledge for reasoning
• Combining Both - Customer support agent using memory + knowledge
• OpenAI Agents Library - Function tools and agent orchestration
• Production-Ready Patterns - Async/await, proper initialization, tool design

greatminds-demo/
├── demos/
│   ├── 01-memory-graph/       # Graphiti + FalkorDB for conversation memory
│   ├── 02-knowledge-graph/    # FalkorDB + Cypher for domain knowledge
│   └── 03-multi-agent/        # Combined: memory + knowledge graphs
└── README.md                  # This file

• Python 3.9+
• Docker (for FalkorDB)
• OpenAI API key

Bureaucratic interaction tracker that learns domain dynamically through conversation

Use Case: Managing ongoing cases with Israeli government agencies and service providers (tax authority, insurance, healthcare, utilities) - tracking state, deadlines, promises, and requirements without predefined ontology.

Stack:

• Graphiti for dynamic episodic memory (learns structure from conversation)
• FalkorDB as graph backend
• OpenAI Agents library (GPT-4)
• FastAPI server

Key Features:

• Tracks multiple parallel cases across organizations
• Learns domain structure dynamically (no fixed schema)
• Maintains state: Open → Submitted → Resolved
• Semantic search: "What am I waiting for?", "Show tax case status"
• Temporal reasoning: deadlines, durations, sequences
• Demonstrates WHY graph beats plain chat history

Start Demo:

cd demos/01-memory-graph
cp .env.example .env  # Add OPENAI_API_KEY
./start.sh

Try Asking:

• "I opened case #12345 with Tax Authority for refund"
• "What's the status of my tax case?"
• "What am I waiting for right now?"

Suspicious identity and Digital Footprint Analysis

Stack:

• FalkorDB for knowledge storage (port 6379)
• YAML ontology definition
• OpenAI Agents library
• Cypher query generation

Key Features:

• Two-step query process (schema → generate query → execute → format)
• Externalized ontology in YAML
• Tool logging with 🔧 markers

Start Demo:

cd demos/02-knowledge-graph
cp .env.example .env  # Add OPENAI_API_KEY
./start.sh

Try Asking:

• "Which profiles share the same email address?""
• "Show me profiles using disposable emails"
• "Which profiles are accessing from VPN addresses?"

Support chatbot combining both graph systems

Stack:

• FalkorDB (port 6380) for both graphs
• Graphiti for session memory
• Knowledge graph for problems/solutions
• OpenAI Agents library
• Unified agent with 9 tools

Key Features:

• Memory Tools (5 async): start_session, record_issue, record_step, get_session_context, search_similar_past_issues
• Knowledge Tools (4 sync): get_knowledge_schema, search_problems, get_solutions, execute_knowledge_query
• Session tracking with ContextVar
• Customer support domain (auth, payment, email, database systems)

Start Demo:

cd demos/03-multi-agent
cp .env.example .env  # Add OPENAI_API_KEY
./start.sh

Memory Graphs (Graphiti + FalkorDB):

• ✅ Episodic memory for conversations
• ✅ Automatic entity/relationship extraction
• ✅ Semantic search across history
• ✅ Time-aware context retrieval

Knowledge Graphs (FalkorDB):

• ✅ Structured domain knowledge
• ✅ Deterministic reasoning via Cypher
• ✅ Complex graph traversal queries
• ✅ Consistent, explainable answers

Combining Both:

• ✅ Knowledge provides expertise
• ✅ Memory provides context
• ✅ Agent orchestrates between both
• ✅ Powerful for support, research, analysis

MIT License - feel free to use these demos for learning and building your own projects.

Questions? Open an issue or reach out!

Talk: LevTech GreatMinds #10
Topic: From Prompt Parrots to Real Agents
Date: Dec 10th, 2025