Graphiti

Time is the organizing idea in Graphiti, which builds knowledge graphs for agent memory that record not just facts but when each fact became true and when it was superseded, using a bi-temporal model that tracks event time and ingestion time separately. Instead of batch recomputation, new episodes are integrated incrementally in real time: an LLM extracts entities and relationships, resolves them against the existing graph, and invalidates outdated edges rather than deleting history, so an agent can answer what was true at a given moment. Retrieval combines semantic embeddings, BM25 keyword search, and graph traversal without requiring an LLM call at query time, keeping lookups fast. The Python library is Apache-2.0 licensed, runs against Neo4j or FalkorDB among other backends, supports OpenAI, Anthropic, Gemini, and local models for extraction, and underpins Zep's commercial memory platform, with the architecture described in an arXiv paper. Roughly 29,000 GitHub stars reflect its position as the leading open temporal graph memory layer for agents.