Provenance Is the Missing Layer in LLM Agent Memory

LLMs destroy the paper trail when synthesizing facts from multiple sources. A confident allergy fact may hide an unverified patient self-report, risking clinical errors. Post-hoc source tagging fails because facts mutate, are derived from many inputs, and merge. Only a graph linking every derived artifact to immutable source episodes survives.

Graffiti, an open-source temporal graph framework, keeps sources verbatim and chains facts as triples to them. When entities merge or facts are invalidated by new data, the graph topology captures the full lineage. Metadata projection allows tagging sources (e.g., EHR-verified) so facts inherit veracity flags, enabling policy-driven retrieval for life-or-death prescriptions.

Deletion becomes equally precise: if a source must be removed for compliance, facts supported by remaining sources survive; those with no remaining support are automatically deletable. This avoids over-deletion while honoring retention rules. The approach turns compliance from an operational nightmare into a structural property.

The major tradeoff is cost. Graph construction and lineage maintenance are expensive. The team reduced latency and token spend by mixing LLM calls with deterministic NLP, sim hash, and entropy-based deduplication. For engineering leaders, provenance must be engineered into the data model from day one, especially in multi-agent, multi-source settings.