Architecture and analysis preprint
Consolidation Leakage and Lineage-Closure Projection in Shared Multi-Agent Memory
Marvin Danig · 2026-11-20
Abstract
Shared memory for teams of people and agents derives new memories from old ones: summaries, semantic facts, and learned procedures. Access control usually checks only the record being read. A derived memory can depend on sources the reader may not see, so it can pass restricted information to that reader even though no restricted record is read. We call this consolidation leakage and define it against a counterfactual memory built without the restricted sources. We describe lineage-closure projection, a filed design that checks a memory's ancestry as well as the memory itself, and analyze what it prevents and what its redactions still disclose. We also compute how often small-sample evidence gates promote unreliable learned procedures. The claim is falsified if record-level filtering shows no leakage above the counterfactual baseline, or narrowed if per-scope consolidation matches projection at equal cost.
Research boundary: this is an architecture-and-analysis preprint, not a benchmark report. Its only numbers are exact binomial calculations. It defines claim boundaries, falsification tests, and an evaluation protocol for future measurement. Copyright 2026 Achiral Research. Patent rights reserved. No patent license is granted by publication of this article.
Topics
Canonical page: https://achiral.ai/papers/consolidation-leakage-and-lineage-closure-projection-in-shared-multi-agent-memory