Citation Needed: Provenance for LLM-Built Knowledge Graphs — Daniel Chalef, Zep AI

An AI memory needs to retain where a statement came from even after new information changes it. In Zep AI's provenance talk, its team describes linking extracted facts to the original records that support them. In the company's Graphiti framework, merging two representations of the same person should preserve both sets of source links.

Those connections make updates and deletion more precise. The talk describes retaining a fact when one source is deleted but other sources still support it, and removing a fact whose only supporting source is gone. Applications can inspect that history to decide which evidence to trust. Knowing a statement's origin enables that decision; it does not by itself prove the statement true.

“Synthesis often destroys the paper trail of how these outputs were originated.”

Daniel Chalef

“So provenence in a context store containing facts is a knowledge graph.”

Daniel Chalef

“Lineage needs to be built in to the data structure, engineered into the data structure, which is a graph, not logged afterwards.”

Daniel Chalef

“A fact is only deleted if no remaining episodes support it.”

Daniel Chalef

“markdown suffers from provenence. File-based um file-based uh memory starts to break down with provenence.”

Daniel Chalef