getzep/graphiti

Graphiti is a Python library for giving AI agents a long-term memory that understands time. Instead of an AI agent that forgets everything between conversations, Graphiti stores facts as a knowledge graph, a web of interconnected pieces of information, where each fact is tagged with when it was true and when (if ever) it changed or was superseded. The practical problem it solves: if your AI assistant learns on Monday that "the user's favorite restaurant is Noodle House," and on Friday the user says "I stopped going to Noodle House," a naive system might still recommend it. Graphiti tracks these temporal changes, it knows what is true now versus what used to be true, making the agent's memory accurate over time rather than just a pile of accumulated text. Under the hood, it builds a graph database of entities (people, products, concepts), relationships between them, and the time windows during which each relationship was valid. When you query it, it combines semantic search (meaning-based), keyword search, and graph traversal to find the most relevant and current information quickly. For a vibe coder building an AI chatbot, personal assistant, or agent: if your users interact with your AI repeatedly over time and you want it to remember and reason about evolving context (their preferences, history, changing situations), Graphiti handles the memory layer. You connect it to Neo4j (a graph database) and integrate it into your agent with a few lines of Python. The commercial version is called Zep and handles scaling, deployment, and operations, Graphiti is the open-source core. Both are backed by research published on arXiv.