- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 1
- Soundness
- 1
Read review
Paper summary
This paper proposes lifecycle authorization for a personal agent’s persistent memory. Records inherit their origin audience; derived records are partitioned, restricted by audience intersection, explicitly declassified, or suppressed. Object-specific grants remain separate, and admission is checked against a snapshot before each physical model attempt. The policy is implemented in two independently persisted reference architectures and descriptively integrated into a native agent runtime.
The conditional policy appears sound, and exact assembled contexts are an appropriate endpoint for the stated confidentiality property. The 10,000-history confirmation provides credible implementation and retrieval evidence within its controlled generator. However, the strongest results mainly exercise a familiar admission predicate, while the distinguishing lifecycle contribution receives substantially thinner validation. My recommendation is marginally below acceptance because the paper has not yet established enough additional capability beyond existing authorization and information-flow mechanisms.
Critical strengths
Precise and implementable policy semantics.The contract clearly distinguishes origin audiences, separate grants, transitive source currentness, and per-attempt snapshots. In particular, separate grants cannot be unioned to authorize a larger audience, and source grants do not automatically authorize derived records.
An appropriate structural security endpoint.Measuring forbidden items in the exact assembled context avoids relying on the model to suppress information after reading it. The paper correctly separates this property from answer-level leakage and broader contextual integrity.
Meaningful implementation validation.B-FLAT and B-GRAPH use independent schemas and query paths. Differential, mutation, and transformation tests are useful evidence that the policy is implemented across multiple access pat
Critical weaknesses
- The research increment over established information-flow control needs stronger demonstration
Audience containment, conservative propagation of derived-data permissions, and explicit declassification are established information-flow concepts. The paper acknowledges this foundation, but its discussion of agent IFC emphasizes control and effect flows more than the confidentiality mechanisms already present in that literature.
For example, FIDES explicitly describes permitted-reader sets, intersection for derived data, email recipients as a source of confidentiality labels, and conservative propagation through LLM histories and outputs. These overlap substantially with the paper’s basic labeling mechanisms. FIDES, Sections 4.1–4.3.
The defensible increment is therefore the persistent-memory lifecycle integration: exact grants, source currentness, durable projections, and admission for each physical attempt. This is potentially valuable engineering, but the paper needs to show more directly what that integration enables or prevents beyond a conventional audience-aware store. The conditional proofs largely establish faithful enforcement of the specified policy; they do not by themselves demonstrate a substantial new capability.
- The most distinctive lifecycle mechanisms receive weaker evidence than the inherited admission rule
The large confirmation establishes context exclusion, recall parity, and resistance to unauthorized distractors. Direct evidence for derived-memory handling consists mainly of casewise transformation tests, eight intersection-versus-suppression cases, and descriptive native scenarios.
Importantly, Appendix E.2 explicitly states that grant revocation, item deletion/revocation, and member removal are not evaluated. The full conformance counts reported in Section 6.1 also come from an earlier revision, while the evaluated revision has a smaller selected check set.
The native results do not fully close this gap. They span heterogeneous earlier revisions, and approved widening and complete intersection each produce five incorrect abstentions. Incomplete-provenance scenarios record non-emitting requests rather than demonstrated runtime refusals (Appendix E.6).
These observations do not refute the conditional guarantee. They limit the evidence for the paper’s central extension across the memory lifecycle. A focused replay on one frozen revision should follow derived artifacts through grants, source revocation or membership changes, persistence/restart, and the next model attempt. Inputs, source closures, resulting labels, and assembled contexts would make this contribution directly assessable.
- The recall study establishes a narrow operating regime rather than a general absence of recall cost
Both architectures match principal postings and bounded post-filtering exactly in every history and at every reported (k). This is positive evidence for policy conformance and the stated non-inferiority objective.
However, the post-filter’s 20-item window is never exhausted by the eight-distractor manipulation. The main superiority result therefore compares authorized retrieval with unscoped retrieval under deliberately introduced unauthorized competitors; it does not distinguish the proposed lifecycle implementation from the strong authorization-aware alternatives.
The independently authored corpora do not resolve this limitation. Their audience overlay contains no competing cross-audience evidence, so all arms—including unscoped retrieval—produce identical retrieval results (Appendix E.4).
Consequently, the unqualified statement that authorization before ranking “costs no recall” should be restricted to the evaluated lexical retrieval regime. One additional regime with a binding candidate window, greater audience overlap, or nontrivial derived-memory competition would be more informative than further repetitions of the current generator.
- The semantic outcome claims lack sufficiently independent validation
The primary context-exposure and retrieval measurements are deterministic and do not depend on an LLM judge. The answer-level evidence is less secure.
All 15,229 semantic-review subjects receive AI judgments through several routes, none is independently validated, and the approved consistency audit was not executed (Appendix E.5). Native correctness labels were assigned by the executing AI session. With only seven detected substitutions in the unscoped arm, modest classification error could materially affect the observed counts.
The paper appropriately does not claim that H3a/b succeeded. Nevertheless, “zero substitutions” should consistently be described as zero detected substitutions under the current labeling procedure. Blinded verification of the seven positive cases and their paired architecture answers, together with an independently reviewed negative sample, would strengthen the semantic interpretation without changing the prespecified statistical decision.
Questions for authors
What lifecycle behavior is demonstrated beyond an audience-aware store with conventional label propagation and object-specific grants? Can this be isolated against the previous single-rule implementation?
Can the final evaluated revision be replayed through complete derivation–grant–revocation–restart trajectories, including the four retained unfavorable native scenarios?
How do recall and candidate utilization behave when the post-filter window becomes binding or derived items compete with authorized raw evidence?
Can the rare substitution labels and a stratified sample of negative judgments receive blinded independent verification?
Response to Reviewer xpEq
Thank you for the detailed review.
Increment and isolation. The contribution is persistent-memory lifecycle integration; §2 explicitly credits FIDES's reader sets and intersection. On the fixed panel (Table 7), the complete mechanism exposes 0/27 opportunities per reference store and 0/8 natively. Removing refresh exposes 19/27 and 6/8 respectively; retrieval-only checking exposes 5/8 natively. Derived labels retain 18/18 challenge and 16/16 ordinary operations per reference store, versus 0/18 and 0/16 under suppression. These are controlled omissions within our implementation, not a head-to-head run of the earlier system. Separate evidence comes from Shen et al. (2026): default Graphiti retrieval returns revoked policies (App. D).
One-version trajectory and native scenarios. One post-submission candidate passes 12 controls using persisted SQLite authorization and native admission into a local receiver. Source invalidation blocks a granted derivative after restart; an unaffected item remains usable, and incomplete source records are refused (App. F.4). Both stores and the retrieval, context-assembly, grant-authority, oracle and provenance modules match the submitted artifact byte-for-byte (supplement,
CODE-COMPARISON.json). All four native scenarios are replayed from retained candidates. The current query-aware tie-break selects admissible evidence for approved widening and complete intersection. Overlapping audiences remains a top-1 miss; consolidation still misses some supporting evidence.Recall regime. Parity is now scoped to complete candidate coverage (§6.2). The added LoCoMo regime, with constructed audiences, makes the window bind: a 20-slot post-filter underfills 145/1,981 queries. Recall@5 over 1,531 eligible questions is 56.03% with that window versus 56.39% authorizing first (Table 22).
Human verification. Yes. A non-author reporting no prior exposure assessed 141 cases with method names and prior verdicts hidden: seven flags, 14 paired answers and 120 stratified random negatives. An author with prior exposure assessed 57 overlapping cases. Each originally marked two flags as substitutions, sharing one positive. Neither marked a definite substitution in the paired answers. The non-author's negatives contain 0 Yes, 117 No and 3 Not sure. Original and corrected records remain separate (App. F.8, Table 24).
Presentation. The prose is simplified; the formal model, proofs, hypotheses and main result tables are retained. The general response lists what changed.