Graphiti solves a fundamental limitation of traditional RAG systems by providing real-time incremental knowledge graph construction instead of batch processing. When new information arrives as episodes of text, JSON, or chat messages, Graphiti extracts entities and relationships, resolves them against existing graph nodes through a three-tier deduplication strategy combining exact match, fuzzy similarity, and LLM reasoning, and detects contradictions that trigger temporal invalidation of outdated facts.
The bi-temporal data model is what sets Graphiti apart from every other knowledge graph framework. Every edge carries explicit validity intervals tracking both when an event occurred and when it was recorded. This enables powerful historical queries where agents can reconstruct the state of knowledge at any point in time. Combined with full provenance tracing from derived facts back to source episodes, Graphiti provides the auditability that enterprise applications require.
Graphiti supports multiple graph backends including Neo4j, FalkorDB, Kuzu, and Amazon Neptune, with LLMs spanning OpenAI, Anthropic, Gemini, and Groq. The MCP server lets Claude, Cursor, and other assistants interact directly with knowledge graphs. The framework powers Zep's commercial platform and demonstrates state-of-the-art agent memory performance, outperforming MemGPT with 94.8% accuracy on the Deep Memory Retrieval benchmark.
