Lineage-Aware Memory Governance: A Derivation-Gated Framework for Privacy-Preserving Column-Level Access Control in Enterprise AI Agents
The paper proposes gating shared agent memory by the full column lineage behind each cached result.
Its Analytical Memory Unit attaches a derivation graph to cached outputs and only returns them when the requester is authorized for every source column involved. The authors say naive content-gated memory leaked across departments in 18.8-25.5% of cases, while lineage-gated retrieval removed those leaks in six experiments. The guarantee is conditional on complete lineage recording, and the paper treats 90% completeness as a conservative deployment target. A small LLM-SQL proof of concept reported zero leaks over nine round-trips, but the authors frame it as feasibility evidence, not production validation. HF Daily Papers' note
Its Analytical Memory Unit attaches a derivation graph to cached outputs and only returns them when the requester is authorized for every source column involved. The authors say naive content-gated memory leaked across departments in 18.8-25.5% of cases, while lineage-gated retrieval removed those leaks in six experiments. The guarantee is conditional on complete lineage recording, and the paper treats 90% completeness as a conservative deployment target. A small LLM-SQL proof of concept reported zero leaks over nine round-trips, but the authors frame it as feasibility evidence, not production validation. HF Daily Papers' note
score 4