Temporal knowledge graph
A knowledge-graph whose facts carry time, so the store can represent not just what is connected but what is true now vs. what was true before. The defining technique is bi-temporal modeling — each edge records two timelines:
- valid time — when the fact was true in the world (the event occurred);
- transaction time — when the system learned it.
Superseded facts are time-bounded, not deleted, so history is queryable and “what holds now” is unambiguous agent-memory-knowledge-graphs.
Why it matters for agent memory
It closes a gap vector retrieval-augmented-generation structurally cannot: temporal validity. When a user’s situation changes (the canonical example: moving cities), a vector store retrieves the old and new facts as equally relevant — both are semantically close — and the agent can’t tell which is current. A temporal graph orders them, so persistent agent memory reflects the present without losing the past.
Relation to the wiki’s other retrieval gaps
This is the fourth distinct fix in the wiki’s RAG critique, layered on the typed-edge knowledge-graph:
- llm-wiki → no-accumulation gap;
- BM25 (hybrid-retrieval-rag) → exact-token gap;
- typed knowledge-graph (gbrain) → factual-connection gap;
- temporal KG → evolving-facts gap.
It extends gbrain‘s automated, scaled associative-trails with a time axis: the memex trail not only wires associations but knows when each was laid down.
In practice
Built incrementally with graphiti (Zep) over Neo4j, using hybrid retrieval (semantic + BM25
- graph traversal) so the graph carries temporal state on top of the hybrid stack.
Related
knowledge-graph · graphiti · agent-memory-knowledge-graphs · gbrain · retrieval-augmented-generation · associative-trails