The DOI matches its Crossref record; 19 of the 19 references it cites were found in a registry or an index; 11 of the 16 sources its claims cite were checked for support.
Lineage-Aware Memory Governance: A Derivation-Gated Framework for Privacy-Preserving Column-Level Access Control in Enterprise AI Agents
Open accessAbstract
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., MemGPT, Zep, A-MEM) gate retrieval by content, ownership, and role, not derivation, missing a cached insight that embeds a forbidden column. We introduce the Analytical Memory Unit (AMU), a memory schema that attaches a full derivation (lineage) graph to every cached result, gated by a retrieval policy that serves a hit only when the requester is authorised for every column touched. Provided lineage recording is complete, we prove by construction that the policy blocks retrieval of results derived from a sensitive column outside the requester’s permissions, at $O(n)$ worst case – a conditional design guarantee, not an empirical claim, that excludes derived features encoding sensitive information without naming their source. Eliminating measured leakage required 75–90% recorded lineage completeness, so we treat 90% as a conservative deployment target. Across six experiments, lineage-gated retrieval removes the 18.8–25.5% cross-department leakage naive content-gated memory suffers, keeping 81.5–82.6% of memory reuse at $13.8{\,}\mu $ s worst-case overhead. A real-agent proof-of-concept with LLM-generated SQL is consistent with the guarantee: zero leaks over 9 round-trips, two conflicts caught automatically – though a feasibility demonstration, not evidence of production viability. This offers a practical governance layer for shared agent memory, complementing source-layer access control and supporting EU AI Act compliance.
Overview
CoChat Summary
TL;DR
The paper introduces the Analytical Memory Unit (AMU), a shared agent-memory schema that attaches a full derivation/lineage graph to every cached result and gates retrieval on whether the requester is authorised for every sensitive column touched. Under complete lineage recording, it proves a conditional safety guarantee and reports experiments showing lineage-gated retrieval removes an 18.8–25.5% cross-department leakage rate at a 13–14 percentage-point reuse cost and 13.82 µs worst-case retrieval overhead.
Key findings
- Lineage-gated retrieval eliminates the naive content-gated leak rate in both tested schemas: synthetic 5-table leak 18.8 ± 3.7% → 0.0 ± 0.0%, TPC-H 8-table leak 25.5 ± 4.5% → 0.0 ± 0.0%, while retaining 82.6 ± 2.3% and 81.5 ± 4.8% reuse respectively (naive reuse 95.8 ± 0.0% in both).
- The safety guarantee is conditional on complete lineage: at completeness ρ = 0.50, leak rates were 16.3 ± 3.3% (synthetic) and 10.3 ± 4.7% (TPC-H); 0% measured leak required ρ ≥ 0.75 for synthetic and ρ ≥ 0.90 for TPC-H, leading the authors to recommend ρ ≥ 0.9 as a deployment target.
- Runtime overhead is small: gate-check predicate 0.09 µs; retrieve worst case scales from 0.66 µs at n = 1 to 13.82 µs at n = 50, while expected-case retrieve stays near 0.64–0.66 µs; conflict detection is 0.28 µs at n = 1 and 1.63 µs at n = 50.
- In a real-agent proof-of-concept with LLM-generated SQL over Northwind, there were zero leaks across 9 cross-department round-trips, 5 fresh SQL queries, 4 cache-served requests, and 2 automatically detected metric-definition conflicts; the extended 43-pair conflict study found D3 column-graph detector F1 dropped to 0.833 and missed all 8 logic-operator pairs, while D1 exact hash had precision 0.651 and recall 1.000.
Methods
- The core design is formal: Definition 1 specifies the AMU tuple ⟨metric, v, d, τ, L⟩ with lineage graph L; Algorithm 1 gates retrieval on S(a) ⊆ P(d), and Theorem 1 proves sensitive-column safety under Assumption 1 (complete lineage), with corollaries for unsafe store contents and policy updates.
- Experiments use a synthetic discrete-event simulator in pure Python comparing No Memory, Naive Shared Memory, and Lineage-Aware systems; results report mean ± population standard deviation over 30 seeds with bootstrap 95% confidence intervals and Mann-Whitney U tests.
- Six experiments cover a 5-table synthetic schema, an 8-table TPC-H schema, lineage-completeness degradation, a 43-pair fuzzy conflict-detection study, runtime benchmarks over 2,000 iterations at n ∈ {1, 5, 10, 20, 50}, and a real-agent integration using Northwind, LangChain, Ollama, and sqlglot-based automatic SQL lineage extraction.
Limitations
- Stated by the paper: the guarantee holds only under complete lineage recording; measured leakage was eliminated only at 75–90% recorded completeness depending on schema; automatic extraction has blind spots for views, stored procedures, and deep aliasing; inference attacks through undeclared derived features (T4) are not addressed; the 43-pair fuzzy dataset is small; runtime benchmarks come from a single hardware/OS configuration; and the 9-round-trip real-agent result is a feasibility demonstration, not evidence of production viability.
- Stated by the paper: the 18.8–25.5% naive leak rates are properties of the tested workload mixes rather than universal base rates, and the mechanism is one layer in a defence-in-depth stack, not a complete security solution.
- Evident from the text: the main empirical validation is simulator-based and synthetic or benchmark-derived, with no production deployment study; the real-agent experiment is small and limited to SQL query patterns exercised in Northwind, so broader SQL workloads and multi-platform reproducibility remain untested.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
5 Valid · 2 Not valid · 4 Not checked · 72 Own results and interpretations
At least one cited source doesn’t support this claim or points to a different paper, and none confirmed it — see the verdicts below.
In the paper
“Relative to the O (1) dictionary lookup used by all prior systems [4], [5], [6], [7], [8], [9], lineage-gated retrieval trades at most 13 . 82 µ s in absolute terms for the safety guarantee of Theorem 1-a tradeoff Experiments 1-2 show buys elimination of an 18 . 8 -25 . 5% structural leak rate at a 13 -14 percentage-point reuse cost.”
What the sources say
- Disputedcites [4] Packer, 2023, MemGPT: Towards LLMs as operating systems, arXiv:2310.08560 sourceExperimental — accuracy not yet measured
The citation is attached to the claim that prior systems used O(1) dictionary lookup. MemGPT instead describes cosine-similarity vector retrieval using an HNSW index (Section 3.2.1, p. 6) and repeated, paginated retrieval calls (Section 3.2.1, p. 7). It does not establish an O(1) dictionary lookup for MemGPT. (LLM, conf=0.86)
Checked by a reasoning model — may be wrong.
- Disputedcites [5] Rasmussen, 2025, Zep: A temporal knowledge graph architecture for agent memory, arXiv:2501.13956 sourceExperimental — accuracy not yet measured
The citation to Zep is attached to the assertion that all prior systems use O(1) dictionary lookup. Section 3.1 (pp. 4–5) instead describes Zep retrieval using semantic similarity search, BM25 full-text search, and breadth-first graph search, followed by reranking. The 13.82 μs and Theorem 1 claims are not attributed to Zep by this citation. (LLM, conf=0.88)
Checked by a reasoning model — may be wrong.
- Disputedcites [6] sourceExperimental — accuracy not yet measured
The citation is attached to the assertion that prior systems used O(1) dictionary lookup. In §3.4, A-MEM instead describes computing cosine similarity between the query and all stored memory notes, then retrieving the top k. Its Scaling Analysis also reports retrieval time increasing with memory size; the cited paper does not describe its retrieval as O(1) dictionary lookup. (LLM, conf=0.88)
Checked by a reasoning model — may be wrong.
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Disputed
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“at a modest overhead: 4 -8 × storage, a submicrosecond gate check, and at most 13 . 8 µ s worst-case retrieval latency”
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“At ρ =0 . 50 , truncated lineage gives almost no protection (synthetic: 16 . 3 ± 3 . 3% leak; TPC-H: 10 . 3 ± 4 . 7% ).”
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
A conclusion the paper draws from its own results — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“T1, where an agent retrieves a cached result derived from a sensitive column it should not see (mitigated by the S( a ) ⊆ P( d ) gate)”
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: unclear — It is unclear why the paper says it holds.
A conclusion the paper draws from its own results — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“The broader lesson here is that fuzzy-detection evaluations on small, narrow datasets can overstate performance simply by missing edge cases.”
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
16 more claims
Create a free account to see all 83 claims and to verify the 4 references still open.
- It may state two separate claims as one. 51
- Probably not a substantial claim. 22
- The quote only partly states this claim. 1
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
Each contradiction between two passages listed here was confirmed by an independent check.
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ E. EXPERIMENT 6: REAL-AGENT INTEGRATION · p. 9
“Experiments 1-5 assume complete lineage (Assumption 1).”
§ B. EXPERIMENT 3: LINEAGE-COMPLETENESS DEGRADATION · p. 8
“At ρ =0 . 50 , truncated lineage gives almost no protection (synthetic: 16 . 3 ± 3 . 3% leak; TPC-H: 10 . 3 ± 4 . 7% ).”
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ 1 Venkata Sangaraju and 2 Sudhir Vissa · p. 1
“a retrieval policy that serves a hit only when the requester is authorised for every column touched.”
§ A. EXPERIMENTS 1-2: SYNTHETIC AND TPC-H SCHEMAS · p. 7
“The gate enforces restrictions only on columns in S ; non-sensitive blocked columns represent full policy intent but are not currently gateenforced.”
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ VI. EXPERIMENTS · p. 7
“All experiments report mean ± population standard deviation over 30 seeds, with bootstrap 95% confidence intervals and MannWhitney U significance tests; the full protocol and raw perseed tables are in the Supplementary Material.”
§ E. EXPERIMENT 6: REAL-AGENT INTEGRATION · p. 9
“This is a nine-round-trip proof-of-concept, not a large-scale trial:”
Citations
Cites
The Open Provenance Model core specification (v1.1)
Moreau, L., Clifford, B., Freire, J., Futrelle, J., Gil, Y., Groth, P., Kwasnikowska, N., Miles, S., Missier, P., Myers, J., et al. (2011) · Future Generation Computer Systems · doi:10.1016/j.future.2010.07.005 · Cited by 704 · cited 1×
“The W3C PROV standard [16] and Herschel et al. [17] model data provenance for extract-transform-load (ETL) pipelines, and LINEAGEX [18] extracts column-level SQL lineage with high accuracy.”
A survey on provenance: What for? What form? What from?
Herschel, M., Diestelkämper, R., & Ben Lahmar, H. (2017) · The VLDB Journal · doi:10.1007/s00778-017-0486-1 · Cited by 300 · cited 1×
“The W3C PROV standard [16] and Herschel et al. [17] model data provenance for extract-transform-load (ETL) pipelines, and LINEAGEX [18] extracts column-level SQL lineage with high accuracy.”
AgentLeak: A Benchmark for Internal-Channel Privacy Leakage in Multi-Agent LLM Systems
Yagoubi, F. E., Badu-Marfo, G., & Mallah, R. A. (2026) · IEEE Access · doi:10.1109/access.2026.3704541 · Cited by 83 · cited 2×
“7% of privacy violations [3].”
MemGPT: Towards LLMs as Operating Systems
Packer, C., Wooders, S., Lin, K., Fang, V., Patil, S. G., Stoica, I., & Gonzalez, J. E. (2023) · arXiv · doi:10.48550/arxiv.2310.08560 · Cited by 64
A-Mem: Agentic Memory for LLM Agents
Xu, W., Liang, Z., Mei, K., Gao, H., Tan, J., & Zhang, Y. (2025) · Advances in Neural Information Processing Systems 38 · doi:10.52202/085713-0593 · Cited by 51 · cited 3×
“A-MEM [6] appeared at NeurIPS 2025.”
Simple Prompt Injection Attacks Can Leak Personal Data Observed by LLM Agents During Task ExecutionPreprint
Alizadeh, M., Samei, Z., Stetsenko, D., & Gilardi, F. (2025) · arXiv (Cornell University) · doi:10.48550/arxiv.2506.01055 · Cited by 40
DePLOI: Applying NL2SQL to Synthesize and Audit Database Access ControlPreprint
Subramaniam, P., & Krishnan, S. (2024) · arXiv (Cornell University) · doi:10.48550/arxiv.2402.07332 · Cited by 22
AgentGuardian: Learning Access Control Policies to Govern AI Agent BehaviorPreprint
Abaev, N., Klimov, D., Levinov, G., Mimran, D., Elovici, Y., & Shabtai, A. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2601.10440 · Cited by 18
Zep: A Temporal Knowledge Graph Architecture for Agent Memory
Rasmussen, P., Paliychuk, P., Beauvais, T., Ryan, J., & Chalef, D. (2025) · arXiv · doi:10.48550/arxiv.2501.13956 · Cited by 9
LineageX: A Column Lineage Extraction System for SQL
Zhang, S. H., Miao, Z., & Wang, J. (2025) · Proceedings - International Conference on Data Engineering · doi:10.1109/icde65448.2025.00363 · Cited by 1 · cited 1×
“The W3C PROV standard [16] and Herschel et al. [17] model data provenance for extract-transform-load (ETL) pipelines, and LINEAGEX [18] extracts column-level SQL lineage with high accuracy.”
Research works printed but not linked: 11
- 2026, AI Agent Adoption 2026: 120+ Enterprise Data Points
- 2025, Cost of a Data Breach Report 2025
- 2026, Oracle AI Agent Memory: A Governed, Unified Memory Core for Enterprise AI Agents
- 2024, Why BI Metrics do Not Match Across Dashboards
- 2025, Open Semantic Interchange (OSI): A Vendor-Neutral Semantic Layer Specification for AI and BI
- 2024, Regulation (EU) 2024/1689 Laying Down Harmonised Rules on Artificial Intelligence (EU AI Act)
- 2021, TPC-H Benchmark Specification, Revision 3.0.1
- 2023, SQLglot: SQL Parser and Transpiler
- 2000, Northwind Traders Sample Database
- 2023, LangChain: Building Applications With LLMs Through Composability
- 2024, Ollama: Run LLMs Locally
Cited by
After Attention: The Economics of Markets Where the Buyer Is an AgentPreprint
Vissa, S., & Sangaraju, V. M. (2026) · SSRN Electronic Journal · doi:10.2139/ssrn.7516238
Self-citation
Metadata
- Authors
- Sangaraju, V., & Vissa, S.
- Affiliations
- Independent Researcher
- Georgetown University (US)
- Venue
- IEEE Access
- Published
- 2026
- Volume
- 14
- Pages
- 139683-139693
- ISSN
- 2169-3536 (ISSN-L)
- OpenAlex
- W7207536430
- Semantic Scholar
- 56b1203c7dd05e3f2732242d45dd58324614bb6e
- Cited by
- 1
- References
- 5
- Citations by year
- 2026: 1
- FWCI
- 11.58
- Type
- journal article
- Language
- en
- Open access
- gold
- doi.org
- Locations
- IEEE Access (publishedVersion, cc-by)
- arXiv (Cornell University) (submittedVersion)
- DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion)
- arXiv (Cornell University) (cc-by)
- Topics
- Physical Sciences › Computer Science › Computer Networks and Communications › Access Control and Trust
- Physical Sciences › Computer Science › Artificial Intelligence › Privacy-Preserving Technologies in Data
- Physical Sciences › Computer Science › Information Systems › Digital Rights Management and Security
- Keywords
- data lineage, privacy-preserving retrieval, sensitive data leakage, enterprise AI agents, EU AI Act compliance
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
The claim verdicts this report counts are experimental: their accuracy is not yet measured.
A language model drew the claims from the paper on 2026-10-27, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).
A second model labelled each claim on 2026-10-27 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).
Cited sources are checked with a language model’s help (gpt-6-sol).
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
How to cite
Cite CoChat for the checks on this page, and the paper for its own findings.
- This page
- CoChat. (2026). Citation check: Lineage-Aware Memory Governance: A Derivation-Gated Framework for Privacy-Preserving Column-Level Access Control in Enterprise AI Agents. https://app.cochat.ai/papers/10.1109/access.2026.3730363
- The paper
- Sangaraju, V., & Vissa, S. (2026). Lineage-Aware Memory Governance: A Derivation-Gated Framework for Privacy-Preserving Column-Level Access Control in Enterprise AI Agents. IEEE Access. https://doi.org/10.1109/access.2026.3730363
Data sources
Bibliographic data: Semantic Scholar, OpenAlex, Crossref.
Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).
Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.