// the find
getzep/graphiti
Build Real-Time Knowledge Graphs for AI Agents
Graphiti builds temporal knowledge graphs where facts have validity windows — when something was true, not just what's true now. It sits between raw vector RAG and a full graph database, letting agents query across time, entity relationships, and semantic similarity in one shot. The backing company (Zep) sells a managed version, so OSS Graphiti is the engine without the operational layer.
Bi-temporal tracking is genuinely useful: contradictions invalidate old facts rather than overwriting them, so you can reconstruct agent state at any past point in time. Hybrid retrieval (semantic + BM25 + graph traversal) means you're not gambling everything on embedding quality. The pluggable backend story is real — Neo4j, FalkorDB, Amazon Neptune, and embedded FalkorDB Lite all work via a clean driver interface. The MCP server ships first-class, so wiring this into Claude or Cursor as persistent agent memory is a one-Docker-command setup.
Every episode ingestion calls the LLM multiple times for entity extraction, deduplication, and edge resolution — costs add up fast at any scale, and the default SEMAPHORE_LIMIT=10 cap signals the authors know this bites people. You must run a separate graph database (Neo4j or FalkorDB via Docker) just to get started; there's no in-process option for prototyping that isn't deprecated (Kuzu is being dropped). The OSS version has no built-in performance guarantees — the README explicitly says sub-200ms retrieval is a Zep Cloud feature, not something you get out of the box. Small or local LLMs are a shaky foundation here: structured output reliability is the thing the whole extraction pipeline depends on, and it's fragile below GPT-4 class models.