Lineage-Aware Memory Governance: A Derivation-Gated Framework for Privacy-Preserving Column-Level Access Control in Enterprise AI Agents
A lineage-gated memory policy blocks enterprise AI agents from retrieving results derived from unauthorized columns.
Shared enterprise-agent memory can leak sensitive columns through legitimately computed cached results and can hide conflicting KPI logic. The Analytical Memory Unit stores a full lineage graph on every cached result and returns a hit only when the requester is authorized for every column touched. The authors give a conditional guarantee when lineage recording is complete, and report that eliminating measured leakage required 75–90% completeness. Across six experiments, lineage gating removed the 18.8–25.5% cross-department leakage of naive content-gated memory while retaining 81.5–82.6% reuse at 13.8 microsecond worst-case overhead; an LLM-generated SQL proof of concept showed zero leaks over nine round-trips and caught two conflicts.
- AMU attaches a full derivation graph to every cached result.
- A hit is served only if every source column is authorized.
- Authors treat 90% lineage completeness as a deployment target.
- Gating removes 18.8–25.5% cross-department leakage versus content gates.
- An LLM-SQL proof of concept had zero leaks over nine round-trips.
From the original article
1 min read at huggingface.co
Enterprise AI agents that share a memory store face two unaddressed risks: sensitive data can leak through legitimately computed results the requester could not derive, and departments can silently compute a same-named key performance indicator (KPI) through conflicting logic. Existing agent-memory systems (e. g.
By Venkata M Sangaraju, Sudhir Vissa · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.