The hard part of agent memory is often not retrieving a similar sentence. It is deciding whether that sentence is still true. Graphiti approaches the problem as a temporal context graph: entities, relationships, and facts carry validity windows, and facts can be traced back to the episode that produced them.
That changes the first engineering question. A preference, project configuration, customer relationship, or operational state can be true today and replaced tomorrow. A vector-only retrieval path can still return the old statement because it is semantically close. Graphiti combines semantic retrieval, BM25 keyword matching, and graph traversal, then keeps a path back to the source node used in the result.
I would not treat it as a drop-in memory plugin. The official entry point is pip install graphiti-core. The quickstart also involves a Neo4j or FalkorDB connection, index and constraint initialization, episode ingestion, and hybrid search. A wrong database name, driver URI, LLM endpoint, embedding model, or cross-encoder setting can turn an infrastructure mismatch into a misleading “memory did not update” symptom.
The ontology surface matters too. Graphiti supports custom Pydantic models, while its MCP server includes typed entities such as Requirement, Preference, Procedure, Location, Event, Organization, and Document. That schema is part of the extraction and query contract. It is not a decoration to add after the first demo.
My first verification pass would stay isolated: install the package; prove the Neo4j/FalkorDB connection and database name; add one episode with a known timeline; run a hybrid query; and verify that the result can be traced to its source node and distinguished from a superseded fact. Until write, retrieval, provenance, and contradiction behavior are observable, I would not connect production conversations or a real agent host. The Doramagic pack flags installation and permission risks as high-priority checks for exactly this reason.
This is an independent Doramagic capability pack, not official Graphiti documentation or an endorsement. Project page: https://doramagic.ai/en/projects/graphiti/; Human Manual: https://doramagic.ai/en/projects/graphiti/manual/; upstream: https://github.com/getzep/graphiti.
Top comments (0)