Graphiti
An incremental knowledge-graph ingestion framework from Zep, used to build persistent agent memory as a temporal-knowledge-graph agent-memory-knowledge-graphs. The concrete tooling anchor for this wiki’s temporal-KG thread — the agent-memory counterpart to gbrain‘s self-wiring graph.
What it does
- Incremental ingestion: updates the graph turn-by-turn during a conversation (the demo “visibly builds the graph” as the user talks), rather than batch-rebuilding.
- Bi-temporal modeling: records both when a fact was true and when it was learned, so changing facts are time-bounded instead of overwritten — the core of the temporal-knowledge-graph.
- Hybrid retrieval: combines semantic search, BM25 keyword match, and graph traversal — the same layered design hybrid-retrieval-rag argues is production-correct.
Stack
Typically runs over a Neo4j graph backend, with an LLM (OpenAI in the source) for entity extraction, relationship disambiguation, and embeddings.
Relation to the wiki
Where gbrain auto-wires a typed knowledge-graph for a personal “brain,” Graphiti specializes in the time-aware agent-memory case. Both mechanize associative-trails; Graphiti adds the temporal axis. Cross-spoke: as agent-memory infrastructure it’s adjacent to agentic-tooling-wiki’s harness/memory thread (cf. the gbrain bridge).
Related
temporal-knowledge-graph · agent-memory-knowledge-graphs · knowledge-graph · gbrain · hybrid-retrieval-rag