Kuzu Memory

KuzuMemory

[](https://www.python.org/downloads/) [](https://opensource.org/licenses/MIT) [](https://github.com/kuzu-memory/kuzu-memory/actions)

**Lightweight, embedded graph-based memory system for AI applications**

KuzuMemory provides fast, offline memory capabilities for chatbots and AI systems without requiring LLM calls. It uses pattern matching and local graph storage to remember and recall contextual information.

✨ Key Features

• 🧠 Cognitive Memory Model - Based on human memory psychology (SEMANTIC, PROCEDURAL, EPISODIC, etc.)
• 🚀 No LLM Dependencies - Operates using pattern matching and local NER only
• ⚡ Fast Performance - <3ms memory recall, <8ms memory generation (verified with Kuzu)
• 💾 Embedded Database - Single-file Kuzu graph database
• 🔄 Git-Friendly - Database files <10MB, perfect for version control
• 🔌 Simple API - Just two methods: attach_memories() and generate_memories()
• 🌐 Cross-Platform - Standardized cognitive types shared with TypeScript implementation
• 📱 Offline First - Works completely without internet connection
• 🔧 MCP Ready - Native Claude Desktop integration with async learning support
• 🤖 Hook System - Automatic Claude Code integration using hooks (UserPromptSubmit, Stop)
• 👤 User-Level Memory - Cross-project ~/.kuzu-memory/user.db automatically aggregates your best patterns and rules across all projects
• ⚡ HNSW Vector Search — Kùzu-native approximate nearest-neighbour index (O(log N)) replaces brute-force NumPy cosine scan; embeddings stored at ingestion time
• 🔤 TF-IDF Keyword Boost — multiplicative scoring: final_score = semantic_score × (1 + weight × normalized_tfidf), configurable via KUZU_MEMORY_TFIDF_BOOST_WEIGHT
• **🕸️ Graph Enr