acceptodds
Under review as a conference paper at ICLR 2027

SQCAD: Lifecycle-Aware Authorization for Scope-Qualified Persistent Memory in Long-Term Agents

Abstract

Agent-memory systems typically rank records by their usefulness to the current query, but persistent authorization asks a different question: whether a record may keep influencing decisions that have not arrived. Relevance is calibrated on recorded use, whereas authorization is a claim about decisions that history has not yet settled; refining the first does not by itself answer the second. To address this problem we formalize persistent authorization as a sequential decision problem, first proving that any binary rule reading a score incurs unavoidable worst-case regret when it conflates states whose optimal keep/archive decisions differ, and deriving an exact fixed-continuation decomposition separating representation loss from execution error. Guided by both results, we propose SQCAD, a lifecycle controller whose qualification, authorization, competitive access, and bounded recovery share one scope- and lineage-qualified evidence state. Separating persistent permission from temporary access lets an archived record answer a query without being reauthorized. LifecycleBench makes these decisions measurable by scoring both branches of every keep/archive action under a common continuation. Experiments on the public tasks place the structured-prior guard first: it attains the highest frozen LoCoMo F1 of the evaluated methods at 97.6% less authorized decision memory than the ActMem adapter, and of the LongMemEval-S comparisons that survive Holm correction, 32 of 34 favour SQCAD. Lineage, one retained field, lowers the representation floor by an order of magnitude where it targets it and leaves it unchanged elsewhere. The criterion binds any system that commits a persistent decision on a scalar score, whether the decision governs retention, eviction, or admission. Qualification and lineage raise downstream accuracy alike, so the metrics the field reports cannot tell them apart, highlighting that persistent authorization is a property of the memory decision rather than a by-product of retrieval quality.

Primary area:
foundation or frontier models, including LLMs
Keywords:
agent memory, long-term agents, LLM agents, retrieval, persistent memory, sequential decision making, trustworthy AI, lifecycle authorization, counterfactual evaluation, representation sufficiency

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.25 over 4 reviews, above 49% of the others.
Reviewer_Mp6HOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper frames persistent memory authorization as a keep/archive decision, presents a score-collision lower bound and a fixed-continuation Bayes regret decomposition, and proposes SQCAD, a controller combining scope qualification, lineage, competitive access, and recovery. It evaluates conversational QA on LoCoMo and LongMemEval-S and introduces LifecycleBench, comprising 225 constructed worlds with paired keep/archive continuations.

The distinction between persistent eligibility and temporary retrieval is useful, and the mathematical results appear correct under their stated assumptions. However, I do not think the submission currently establishes a sufficiently strong contribution for acceptance. The certified authorization interface is not instantiated by the evaluated controllers; the public-task improvements do not isolate authorization; and the central lifecycle result largely demonstrates that a rule explicitly representing the generator’s designed conflict can resolve that conflict. These limitations affect the paper’s main argument, rather than only its deployment scope.

Critical strengths
  1. A useful evaluation object. Paired keep/archive continuations make the consequences of a memory decision explicit and allow representation-restricted regret to be distinguished from controller error.

  2. Correct and clearly scoped mathematical identities. The collision lower bound, conditional Bayes decomposition, and refinement result are valid. Appendix A.2 appropriately avoids applying the coarse-summary decomposition to LLM controllers that receive richer information.

  3. Transparent reporting of unfavorable evidence. The paper reports adapter limitations, unchanged physical storage, failed held-out-family transfer, null continuation ablations, and conditional statistical inference. This transparency is a significant strength.

Critical weaknesses
  1. The theoretical results do not establish the advertised certified controller or the necessity of the proposed representation。 Theorem 1 establishes that a representation cannot support uniformly correct decisions after conflating states requiring opposite actions. This is a valid information-loss result, but it is not an intrinsic limitation of scalar representations. For any representation (Z), the scalar (E[\Delta\mid Z]) preserves its binary Bayes decision and information floor. The theorem therefore does not establish that an explicit multi-field state is necessary, nor does it derive the particular SQCAD fields.

Theorem 2 is likewise a standard conditional Bayes decomposition, as the paper acknowledges. Its application to memory decisions is useful, but the theoretical novelty is modest unless it leads to an operational method or a substantial empirical finding.

That operational connection is missing. Section 3.1 and Figure 1 describe a certified interval on the commitment contrast, but the evaluated controllers take qualification from rules or transformed views instead. The paper does not specify an implemented public probe procedure producing appropriately centered contrast observations, or demonstrate interval coverage. The stronger closure conditions in Appendix A.3 are explicitly not estimated or used.

  1. The public-task results do not identify a benefit from persistent authorization LoCoMo and LongMemEval-S evaluate a one-query horizon using contexts constructed before generation. The reported improvements can therefore arise from retrieval, response proposal, smoothing, redundancy control, or packing.

The internal control does not resolve this issue. Table 5 retains the authorization/access machinery while changing the view and budget configuration. Its LoCoMo F1 is approximately 0.117–0.139, compared with 0.263–0.318 for the transformed guards. Appendix B.1 explicitly acknowledges that authorization removal with the view and budgets fixed has not been tested.

This missing comparison is central: the results show that the selected context-construction pipeline helps, but do not establish that persistent authorization is responsible.

  1. The baseline and cost evidence support narrower conclusions than the headline comparison suggests。 The adapters are documented honestly, but several omit capabilities directly relevant to lifecycle control. Trivium omits causal-model revision and its intervention learner; Memory Worth’s end-to-end text transport remains pending; ActMem changes thresholds and retrieval depth; and Oblivion contributes its decay core under a replaced surrounding pipeline (Appendix D).

These are permissible controlled comparisons, but they do not establish that the corresponding memory systems generally lack the distinctions SQCAD introduces. The scalar audit also uses study-defined proxies, with several sharing the same version-update witness; it is not independent evidence of failure across native systems.

The significance count should be interpreted accordingly. Of the 32 favorable LongMemEval-S comparisons, 30 are against the Memory Worth, Trivium, and Oblivion adapters. Only two favor SQCAD over ActMem, both on Terra, while ActMem wins both Qwen comparisons (Table 12).

The 97.6% footprint reduction is correctly defined as authorized decision memory. Its practical value remains insufficiently established: physical storage is unchanged, and LongMemEval-S reader context increases from ActMem’s 746.2 tokens to approximately 1,645–1,714 tokens. A capability-preserving comparator and matched resource trade-off are needed before treating these results as evidence for the proposed authorization design.

Questions for authors

Which evaluated controller implements the certified interval, what does a probe observe, and what empirical coverage or sign-error guarantee does it satisfy?

With candidates, FIR/SP, guard access, packing, and budgets held fixed, what changes when persistent authorization is disabled?

Does the lineage result survive independently constructed mechanisms or changes to the utility ratios? How much of the remaining information floor is attributable to unavailable future information?

Can the authors provide a matched retrieval-only comparator and a temporal-maintenance comparator retaining the capabilities relevant to this task?

1 comment
AuthorsOpenReview ↗

Certified interval, utility sensitivity, and temporal baselines

We thank the reviewer for the theory reconstruction and the scalar counterexample. We agree that the original paper left the certified interval disconnected from an evaluated controller, did not isolate authorization, and lacked code-reproducible comparators.

W1 — The theoretical results do not establish necessity of the particular representation

Theorem 1 is a bound on a scalar that conflates states with opposite optimal actions; for any Z, E[Δ|Z] preserves the binary Bayes decision and the information floor; the theorem does not establish necessity of the five fields. The revision states that E[Δ|Z] is evaluator-side. The necessity claim is about the contrast partition: on an unidentified pair the best committing rule has error at least 1/2 and regret at least m₊m₋/(m₊+m₋). The theorem constrains rules that must choose under the same visible information; whether an additional probe or abstention lowers regret depends on the information it yields and its cost.

W2 — Which evaluated controller implements the certified interval, and what does a probe observe?

The evaluated controllers do not construct the interval. The evaluator generates noisy observations around a training-cell mean and tests the interval against held-out branch contrasts. At the nominal 95% level, the lineage interval covers 86.8% in family and 60.7% under family holdout. Table 7: in-family 0.868 [0.776, 0.937], wrong-sign ≤0.001 (fraction of all 225 worlds, not commits), abstention 0.595; family-holdout 0.607 [0.298, 0.841], wrong-sign 0.087. The gated controller gives 0.8953 vs 0.9443 (improvement 0.049 [0.000, 0.146], lower end zero) and fails under family holdout.

The simulated observations are centered on a training-cell mean, while coverage is scored against each held-out world's branch contrast. Switching to residual resampling produced similar undercoverage, so changing the noise model alone did not resolve this mismatch. Table 7 reports the average per-draw commit probability (0.405 at five observations, hence abstention 0.595); the gate commits when a majority of its 100 draws commits, yielding a 0.369 world-level commit rate. The two rates use different aggregation units and need not coincide.

W3 — Public-task results do not identify persistent authorization

Both public datasets have a one-query horizon, so they cannot identify the long-horizon value of authorization. We add the fixed-view, fixed-packing, fixed-budget ablation; auth-off keeps every visible message authorized, so the eligible set grows and this is not a literal fixed-pool comparison. We repeated the paired comparison at the original per-model answer caps (512, or 1,024 for GLM). All five on-minus-off point estimates remained negative, and only DeepSeek's paired interval excluded zero. Absolute F1 levels did not fully reproduce across runs; the DeepSeek identifier also differed, so we interpret the within-run contrasts and leave the remaining cross-run differences unattributed.

On full LoCoMo the four readers give −0.0320 to −0.0388 with intervals excluding zero. At a matched 12-item budget the full-data differences are about 0.019 F1: LongMemEval-S is a tie for both controls, while on LoCoMo both random and relevance selection are slightly better than lifecycle selection. The paired controls provide no evidence of a QA gain from enabling authorization.

At the fixed 12-item budget the guards authorize 323.9/330.3 tokens and the untransformed control 329.9; the 97.6% refers to the guards' authorized tokens relative to ActMem's full-store 13,343.8. Of the 32 SQCAD-favoring comparisons, 30 are against Memory Worth/Trivium/Oblivion; the two against ActMem are on Terra.

W4 — Baseline and cost evidence support a narrower conclusion

Utility sensitivity, residual uncertainty, and temporal baselines. Appendix B.3 replays frozen actions while varying task value, harm, exposure, probe cost, and storage cost one at a time. Across the tested settings, lineage incurs 8.8–9.8 lower regret than qualification. The harm sweep is uninformative because these frozen actions incur no wrong-use penalty. We have not separately identified the fraction of the residual 0.0634 floor attributable to unavailable future information. The 12-item relevance-only control addresses the retrieval comparison. A capability-preserving temporal-maintenance comparison remains open.

On LongMemEval-S, the reader context rises from 746.2 tokens for the ActMem adapter to 1,644.8–1,713.6 for the SQCAD guards, while physical storage is unchanged.

The displayed rows are matched-contract adapters, not native systems; Table 11 separates evidence tiers. Full LoCoMo implementation checks: A-MEM 0.2710, MemoryOS 0.2266, Mem0 0.0551/0.0445 (Table 12). We do not claim QA SOTA.

Where to find the changes

Tables 3, 6–7 and 11–12, and Appendices A–E; the sections above name the relevant table for each point.

Reviewer_8gzQOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies persistent memory authorization in long-term agents, distinguishing whether a memory should continue influencing future decisions from whether it should be retrieved for a current query. It introduces a conditional lower bound for score-based binary decisions, a regret decomposition separating representation loss from execution error, and SQCAD, a lifecycle controller combining qualified evidence, lineage, and recoverable archives. The authors also construct LifecycleBench with 225 paired keep/archive worlds. Experiments show memory-efficiency and QA results, while lineage reduces the empirical representation-loss floor. However, its advantage does not persist in held-out-family evaluation.

Critical strengths
  • Meaningful problem formulation. The distinction between persistent authorization and query-time retrieval is well motivated. Allowing archived memories to be temporarily accessed without restoring persistent permission provides a useful framework for long-term memory management.

  • Clear theoretical framework and diagnostic evaluation. The conditional lower bound and regret decomposition are mathematically sound and help distinguish representation limitations from controller errors. LifecycleBench's paired counterfactual outcomes make these quantities measurable within the constructed environment.

  • Promising empirical results and transparent reporting. SQCAD achieves competitive or improved QA performance with substantially less authorized memory. The paper also reports unfavorable results, negative ablations, baseline implementation differences, and statistical limitations, supporting a balanced interpretation.

Critical weaknesses
  • The proposed certification mechanism is not empirically validated. Section 3.1 motivates authorization through confidence intervals over keep/archive value differences, but the evaluated controllers rely on heuristic qualifications or LLM decisions rather than constructing these intervals. The paper does not establish how such contrasts can be reliably estimated from online observations or verify the required coverage assumptions. This leaves a gap between the theoretical framework and the implemented system. An empirical evaluation of interval coverage and wrong-sign decisions would strengthen the contribution.

  • The contribution of persistent authorization to downstream performance is not isolated. The public-task experiments jointly change retrieval views, packing, and authorization behavior. Consequently, the reported QA improvements may primarily result from FIR/SP retrieval transformations rather than lifecycle authorization itself. A controlled authorization-on/off ablation with identical retrieval views, candidate pools, and budgets is needed to establish the contribution of the central mechanism.

  • The demonstrated representation gains have limited generalization evidence. LifecycleBench consists of constructed worlds, and version updates account for 89.8% of the qualification policy's regret. Lineage directly addresses the explicitly represented conflicts in this mechanism. Moreover, the fitted-summary advantage disappears under held-out-family evaluation (Table 14). While this does not invalidate lineage's usefulness, additional evaluation on independently designed scenarios, ambiguous updates, and coexisting memory versions is needed to support broader generalization claims.

  • The theoretical novelty and implications could be more clearly positioned. Theorem 1 follows a conditional indistinguishability argument, while Theorem 2 is closely related to standard binary Bayes-risk decomposition, as the paper acknowledges. Neither establishes the necessity or sufficiency of SQCAD's particular representation fields. Furthermore, the scalar lower bound applies only when a score conflates states with opposing optimal actions. The paper should clarify that its primary novelty lies in the authorization formulation and evaluation framework, and avoid implying a general limitation of scalar representations.

Questions for authors
  1. Certification vs. implementation: Is the confidence-interval-based qualification proposed in Section 3.1 actually implemented in any evaluated controller? If not, how should readers interpret the theoretical certification guarantees in relation to the reported empirical results?

  2. Contribution of authorization: Since the public-task experiments jointly modify retrieval views and authorization behavior, what evidence supports attributing the observed QA and memory-efficiency improvements specifically to persistent authorization rather than FIR/SP retrieval transformations?

  3. Generalization of lineage: Given that version updates dominate the observed regret reduction and the fitted-summary advantage disappears under held-out-family evaluation (Table 14), what evidence supports the generalizability of lineage-based authorization beyond the explicitly constructed version-update scenarios?

  4. Theoretical novelty: Given the connection of Theorems 1 and 2 to standard binary decision theory, what is the main theoretical advancement beyond applying existing decision-theoretic results to memory authorization? Do the results establish any necessity or sufficiency of SQCAD's proposed representation fields?

3 comments
AuthorsOpenReview ↗

Certificate coverage and the fixed-view authorization control

We thank the reviewer for the structured critique. We agree with the four concerns and address each where the evidence allows.

W1 — The certification mechanism was not empirically validated

The evaluated public-task controllers do not construct the interval. The evaluator generates noisy observations around a training-cell mean and tests the interval against held-out branch contrasts. At the nominal 95% level, the lineage interval covers 86.8% in family and 60.7% under family holdout, so it undercovers. Table 7 gives the full diagnostic: in-family 0.868 [0.776, 0.937], wrong-sign ≤0.001 (fraction of all 225 worlds, not commits), abstention 0.595; family-holdout 0.607 [0.298, 0.841], wrong-sign 0.087. The gated controller gives 0.8953 vs 0.9443 (improvement 0.049 [0.000, 0.146], lower end zero), with a 0.369 commit rate and zero wrong-sign among commits; it fails under family holdout.

The simulated observations are centered on a training-cell mean, while coverage is scored against each held-out world's branch contrast. Switching to residual resampling produced similar undercoverage, so changing the noise model alone did not resolve this mismatch. We have no measured coverage guarantee for a deployed probe; the closure conditions in Appendix A.3 are not estimated.

Table 7 reports the average per-draw commit probability (0.405 at five observations, hence abstention 0.595); the gate commits when a majority of its 100 draws commits, yielding a 0.369 world-level commit rate. The two rates use different aggregation units and need not coincide.

W2 — The contribution of persistent authorization was not isolated

We add the fixed-view, fixed-packing, fixed-budget authorization-on/off ablation. The view, candidate rule, packing and item budget are fixed; auth-off keeps every visible message authorized, so the eligible set grows and this is not a literal fixed-pool comparison.

On full 500-trace LongMemEval-S, all five full-run point estimates favor auth-off; only DeepSeek's interval excludes zero (−0.0197 [−0.0367, −0.0029]). We repeated the paired comparison at the original per-model answer caps (512 tokens, or 1,024 for GLM). All five on-minus-off point estimates remained negative, and only DeepSeek's paired interval excluded zero. Absolute F1 levels did not fully reproduce across runs. The DeepSeek identifier also differed, so we interpret the within-run contrasts and leave the remaining cross-run differences unattributed.

On full LoCoMo (1,986 QA), the four readers give −0.0320 to −0.0388, with every conversation-cluster interval excluding zero. At a matched 12-item budget the full-data differences are about 0.019 F1 and do not consistently favor authorization: LongMemEval-S is a tie for both controls, while on LoCoMo both random and relevance selection are slightly better than lifecycle selection. We no longer claim an authorization-driven QA gain.

W3 — The representation gains have limited generalization evidence

We agree. The in-family 1.7196 → 0.0634 result (Table 3) and the family-shift 1.9851 result (Table 16) answer different questions; Table 17 gives 0.0713 when the version-update family receives zero weight. R3 reduces qualification regret from 118.68 to 0 on version_update/update_before and leaves other families unchanged. A witness pair shares the overlap+scope+resolved summary, with τ = +43.45 and −134.17. The 78 LongMemEval knowledge-update questions give 0.4639 (archive old), 0.2645 (keep old before), and 0.3593 (keep old after); the keep contexts are longer, so this is not a pure content intervention. None of these is an independent mechanism family.

W4 — Theoretical novelty and implications

We clarify that Theorem 1 constrains a scalar that conflates states with opposing optimal actions, not every scalar representation. For any Z, E[Δ|Z] preserves the binary Bayes decision and the information floor; the theorem does not establish necessity of the five fields. E[Δ|Z] is evaluator-side. The necessity claim is about the contrast partition: on an unidentified pair the best committing rule has error at least 1/2 and regret at least m₊m₋/(m₊+m₋). The theorem constrains rules that must choose under the same visible information; whether an additional probe or abstention lowers regret depends on the information it yields and its cost.

Where to find the changes

  • Table 7 and Appendix B.3: coverage, wrong-sign, abstention, the gated controller, and the residual bootstrap.
  • Table 6 and Appendix B.2: full 500-trace and full LoCoMo ablations plus matched-budget controls.
  • Table 3, Table 16 and Table 17; Appendix A.2: decomposition, family shift, family weights, R3, and the witness pair.
  • §3.1 and Appendix A.1: Theorem 1 scope, evaluator-side E[Δ|Z], and gating necessity.

I appreciate the authors' transparency in reporting negative results and clarifying the scope of their claims.

The response addresses my concern regarding the interpretation of the theoretical results (W4). However, the main empirical concerns remain:

  • Certification (W1): The substantial undercoverage, particularly under family holdout, indicates that the proposed certification mechanism remains empirically unvalidated.
  • Authorization (W2): The new controlled ablations are valuable but show no downstream QA benefit from authorization itself. While QA is not the sole objective, the practical value of the proposed mechanism remains insufficiently demonstrated.
  • Generalization (W3): The lineage gains remain concentrated in constructed version-update scenarios, and the advantage disappears under held-out-family evaluation.

I therefore maintain my original rating.

Thank you for the careful reading, and for acknowledging the transparency of the reporting and the clarification of the theoretical scope (W4).

We agree that the three empirical concerns remain open, and the final version will state them as such rather than soften them. We will report the measured certificate coverage, including the family-holdout case, as an empirical diagnostic with no claim of a coverage guarantee in deployment. We will state that the controlled ablation finds no downstream QA benefit from authorization itself, and withdraw the corresponding claim. We will present the lineage result as a benchmark-specific diagnostic of version-conflict resolution, reported together with the held-out-family outcome.

We appreciate the specificity of the review

Reviewer_Xmq3OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper treats keep/archive decisions in agent memory as a sequential decision problem separate from retrieval. SQCAD keeps a per-memory qualification state and separates persistent authorization from temporary access. In experiments, five LLMs read lexical ranking views under fixed budgets on LoCoMo, LongMemEval-S and LifecycleBench; the qualification rules are tested as deterministic policies.

Critical strengths

LifecycleBench scores both the keep and the archive branch of each decision under one continuation, so retention regret becomes more clearly measurable (Sec. 2.2). Theorem 2 then separates a better rule from a better representation: in Table 4, scope and correction lower only execution error, while lineage lowers the floor from 1.72 to 0.063.

Critical weaknesses

Without its lexical views, SQCAD falls below the strongest baseline, ActMem, on every backbone and both public benchmarks. The view constants have no stated selection protocol (Sec. 3.4), and there is no authorization-removal ablation (App. B.1). The 97.6% authorized-memory saving reflects the item budget versus an ActMem adapter authorizing its whole store; without views the footprint barely moves (329.9 vs 323.9/330.3 LoCoMo tokens, Table 5). Of 32 Holm-significant LongMemEval-S wins, 30 are over adapters scoring at most 0.135 F1 (Table 12).

The qualification state is tested only inside the generator, where the lineage result holds by construction. SQCAD’s LLM controllers on LifecycleBench receive no explicit lineage, validity or censoring field (App. E.2); those rules run only as deterministic policies (Tables 4, 6–7). The generator exposes a same-scope version conflict that lineage detects exactly; this version-update family carries 89.8% of the qualification rule’s regret (Sec. 4.3). With one family held out, lineage’s advantage vanishes: all three summaries reach regret 1.9851 (Table 14). As the theorems are standard decision analysis (p. 4), the contribution rests on this evidence.

Questions for authors

(Clarification) How were the view and packing constants (Sec. 3.4, App. E.2) and the item and reader-token budgets chosen, and were LoCoMo or LongMemEval-S test questions used?

(Clarification) Which experiment supports the abstract’s claim that qualification and lineage both raise downstream accuracy? No LoCoMo or LongMemEval-S run seems to toggle either; Table 7’s agreement column measures decision agreement, not accuracy.

1 comment
AuthorsOpenReview ↗

Narrowing the authorization claim: controls and provenance

We thank the reviewer for the direct diagnosis. We agree that the original paper did not isolate persistent authorization in the public-task experiments, and that the lineage result is tested only inside the constructed generator. The revision adds the fixed-view authorization-removal ablation, removes the unsupported accuracy claim, and narrows the lineage claim.

W1 — Authorization was not isolated; the 97.6% figure needs its budget context

We report the least favorable control first. At the same 12-item budget on full LoCoMo, the Qwen reader scores 0.2425 with lifecycle selection and 0.2619 with relevance-only selection. The paired difference is −0.0193 [−0.0288, −0.0093]. On full LongMemEval-S with DeepSeek, the corresponding difference is −0.0059 [−0.0233, +0.0104]. These controls provide no evidence of a QA benefit from lifecycle selection.

We add a fixed-view, fixed-packing, fixed-budget authorization-on/off ablation. The view, candidate rule, packing and item budget are fixed; auth-off keeps every visible message authorized, so the eligible set grows and this is not a literal fixed-pool comparison. On full 500-trace LongMemEval-S, all five on−off point estimates favor auth-off; only DeepSeek's interval excludes zero (−0.0197 [−0.0367, −0.0029]).

We repeated the paired comparison at the original per-model answer caps (512 tokens, or 1,024 for GLM). All five on-minus-off point estimates remained negative, and only DeepSeek’s paired interval excluded zero. Absolute F1 levels did not fully reproduce across runs. The DeepSeek identifier also differed, so we interpret the within-run contrasts and leave the remaining cross-run differences unattributed.

On full LoCoMo (1,986 QA), the four readers give −0.0320 to −0.0388, and every conversation-cluster interval excludes zero. We do not claim an accuracy gain from authorization.

At the fixed 12-item budget the guards authorize 323.9/330.3 tokens and the untransformed control 329.9. The 97.6% reduction refers to the guards' authorized tokens relative to the ActMem adapter's full-store authorization (13,343.8 tokens); it is an accounting comparison, not a keep/archive effect.

Of the 32 SQCAD-favoring LongMemEval-S comparisons, 30 are against the Memory Worth, Trivium and Oblivion adapters; the two against ActMem are both on Terra. We read this as a component-level comparison, not end-to-end superiority.

W2 — The qualification state was tested only inside the generator

Correct. The released LLM controllers read a view, not the five fields. We removed the LLM comparison from the main tables; Table 15 retains the descriptive pairwise results. The revision calls lineage a benchmark-specific diagnostic. Table 17 gives 0.0713 at zero version-update weight; Table 16 gives 1.9851 for all three summaries under whole-family holdout.

As a boundary test, the R3 unseen-geometry holdout reduces qualification regret from 118.68 to 0 on version_update/update_before and leaves every other family unchanged. This confirms mechanism-specificity, but it is not a cross-mechanism generalization result.

Q1 — How were the view and budget constants chosen?

The FIR kernel (0.25/0.50/0.25), its 0.08 high-pass penalty, the SP offsets (0.10/0.35), the 16-item guard budget and the 12-item workspace budget were fixed in the public contract (commit 98014d9) before the final multi-backbone runs. They were not tuned per backbone or per question. The contract was developed on earlier public-data iterations, and we cannot establish that those iterations were independent of the public evaluation questions, so we do not call the constants pre-registered. Appendix E records this provenance; the Guard-1/2/4 and reader-token variants are sensitivity checks rather than independent validation. Both arms of the auth-on/off ablation use the same constants, so the contrast itself is not confounded by constant selection.

Q2 — The abstract claim that qualification and lineage raise accuracy alike

We removed it. No public-task run toggles qualification or lineage, and the decision-agreement column is decision agreement, not answer accuracy. The abstract now leads with LifecycleBench and treats the public-task results as descriptive.

The public QA comparisons do not show an authorization gain; Table 3's paired-world decomposition separates execution error from representation loss.

Where to find the changes

  • Table 6 and Appendix B.2: full 500-trace and full LoCoMo ablations, plus matched-budget controls.
  • Table 7 and Appendix B.3: coverage and the offline gated controller.
  • Table 3, Table 15, Table 16 and Table 17; Appendix A.2 and Appendix E: decomposition, LLM pairwise results, family shift, family weights, R3 boundary, and constant provenance.
  • Revised abstract and §4.2: accuracy claim removed; 97.6% stated as authorized tokens under the two budget policies.
Reviewer_JoFQOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates persistent memory authorization in long-horizon LLM agents, focusing on sequential keep-or-archive decisions rather than query-time memory retrieval. The authors establish a regret lower bound for scalar-score-based decisions under conflicting action preferences and derive a fixed-continuation regret decomposition that separates representation limitations from controller errors. Building on this formulation, they introduce SQCAD, a state-qualified memory authorization framework, and LifecycleBench, a benchmark designed to assess memory decisions involving scope, lineage, and recovery. The proposed approach is evaluated using established memory question-answering benchmarks alongside paired counterfactual lifecycle scenarios, with experiments examining both downstream task performance and authorization-specific regret.

Critical strengths

1. Well-scoped theoretical contributions with a clear decision-theoretic interpretation.

The theoretical analysis provides a meaningful foundation for understanding persistent memory authorization. Theorem 1 characterizes a fundamental limitation of scalar-score-based binary decisions when the same score corresponds to keep/archive contrasts with opposing signs. Theorem 2 further provides a principled decomposition of Bayes decision regret into representation loss, I(Z)\mathcal{I}(Z), and controller error, E(h;Z)\mathcal{E}(h;Z). An important strength is that these results are explicitly formulated under fixed continuation assumptions, rather than implying that counterfactual branch values can be recovered directly from ordinary retrieval trajectories. This careful distinction strengthens the validity and interpretability of the theoretical contribution.

2. A useful evaluation perspective that separates authorization decisions from retrieval performance.

The paper makes a valuable methodological contribution by distinguishing persistent memory authorization from the quality of query-time retrieval and downstream answering. The experiments illustrate why this separation matters. In Table 3, the SP guard improves LoCoMo performance for Qwen while substantially increasing lifecycle regret, demonstrating that gains in conventional question-answering metrics need not correspond to better authorization decisions. Similarly, Table 4 distinguishes improvements attributable to richer state information from those arising through better use of an existing representation. These findings highlight potential confounding factors in conventional memory evaluations and motivate a more decision-focused assessment of long-horizon memory systems.

Critical weaknesses

1. Limited evidence that the proposed state representation generalizes beyond the constructed benchmark.

A central concern is that the strongest empirical support for the proposed representation is derived from a constructed benchmark whose mechanisms are not independent of the evaluated state features. Table 4 reports a substantial reduction in the representation information floor, from 1.7196 to 0.0634, when lineage information is incorporated across the 225 LifecycleBench worlds. However, Appendix A.2 and Table 14 reveal a significant limitation: under leave-one-mechanism-family-out evaluation, all three representations, including the lineage-enhanced variant, produce the same regret of 1.9851.

The authors acknowledge that the lineage features responsible for resolving version-update conflicts are specific to the corresponding mechanism family and do not encode a decision direction that transfers consistently across families. This concern is reinforced by Table 6, where version-update scenarios contribute 89.8% of the qualification policy's aggregate regret. The observed improvement from lineage therefore appears largely associated with resolving a particular failure mechanism explicitly embedded in the benchmark design.

These findings do not undermine the reported within-benchmark improvements, but they substantially limit the interpretation of those results. In particular, the evidence supports the effectiveness of an explicitly represented version-conflict signal within the constructed environment more strongly than it establishes the general sufficiency or practical superiority of SQCAD's authorization state. A more convincing demonstration would involve independently constructed counterfactual environments or naturally occurring memory trajectories with measurable branch outcomes. Importantly, the scope and lineage representations should be specified before evaluating these additional environments to establish whether the observed benefits transfer beyond the original generator.

2. System-level comparative claims exceed what the matched-contract baseline implementations can establish.

The experimental comparison raises a separate concern regarding the interpretation of baseline results. Although the main performance tables compare SQCAD with several named memory systems, the evaluated baselines are matched-contract adapters rather than complete reproductions of those systems under their original operating conditions.

As documented in Table 10 and Appendix D, the ActMem implementation is a paper-inspired adaptation with modified thresholds and retrieval depth. The Trivium comparison transfers only its probing and logging components, while Memory Worth is evaluated through synthetic calibration consistency without its full text-processing protocol. Oblivion incorporates an official decay component, but its reader and backbone are replaced within the shared experimental setup.

These adaptations may be reasonable for isolating particular mechanisms under controlled conditions. However, they do not directly establish that SQCAD outperforms the original systems in end-to-end operation. Consequently, the broader comparative conclusions should be interpreted as results under a common experimental contract rather than evidence of overall system superiority.

To address this concern, the authors should make the adapter-based nature of these comparisons explicit in the abstract and primary experimental conclusions, restrict performance claims to the evaluated protocol, and, where feasible, include native-system reproductions or stronger comparisons with retention and admission mechanisms operating under their intended configurations.

Questions for authors

Q1. Generalization of the lineage representation.

Table 4 reports a reduction in representation information loss from 1.7196 to 0.0634 after incorporating lineage, whereas Table 14 shows identical regret (1.9851) across all three representations under mechanism-family holdout. How do the authors reconcile these observations when interpreting the general usefulness of lineage? Specifically, should the improvement in Table 4 be understood primarily as a benchmark-specific diagnostic of version-update conflicts, or is there additional evidence that lineage reduces representation loss across previously unseen memory mechanisms? Clarifying the intended scope of this claim would materially affect my assessment of the empirical contribution.

Q2. Scope and interpretation of baseline comparisons.

Could the authors clarify which of the headline performance claims refer to complete baseline systems and which apply only to the matched-contract adapters evaluated in this work? Table 10 identifies substantial modifications to ActMem, Memory Worth, and Trivium relative to their original implementations. Given these differences, it would be helpful to specify whether the comparisons in the abstract and Section 4.2 are intended to establish component-level advantages under a common protocol or broader end-to-end superiority. This distinction is important for assessing the practical significance of the results.

Q3. Numerical consistency in the LongMemEval-S discussion.

Section 4.2 states that the LongMemEval-S guard results are "within 0.038 of the strongest adapter on every backbone." However, Table 2 reports a best SQCAD score of 0.4294 on Terra compared with 0.3139 for ActMem, corresponding to a difference of 0.1155. Could the authors clarify the intended comparison and explain this apparent discrepancy? Resolving it would improve confidence in the quantitative interpretation of the experiments.

3 comments
AuthorsOpenReview ↗

Lineage is a benchmark-specific diagnostic

Thank you again for your comments on our work. Yes. We interpret the lineage result as a benchmark-specific diagnosis of version conflicts. The clearest check is Table 17: when the version-update family receives zero weight, the overlap/scope and lineage representations have the same information floor, 0.0713. The 78 external knowledge-update questions show sensitivity to stale context, but they do not measure paired lifecycle utility.

Q1 / W1 — How the 1.7196 → 0.0634 and 1.9851 results fit together

Table 3 is the in-family decomposition on the same 225 paired worlds: adding scope and correction reduces execution regret from 8.0546 to 7.4966 while leaving the information floor at 1.7196; adding lineage lowers that floor to 0.0634. Table 16 is the family-shift stress test: when a whole mechanism family is held out, all three fitted rules reach 1.9851. The 1.9851 is the regret of a rule fitted without the held-out family, not an information floor, so it is not the 0.0634 floor worsening to 1.9851. These are in-family quantities; the revision does not present Table 3 as a generalization estimate.

We add three boundary checks. The R3 unseen-geometry holdout reduces qualification regret from 118.68 to 0 on version_update/update_before and leaves every other family unchanged. A direct witness shows why: two v3 worlds share the overlap+scope+resolved summary, but update_after has τ = +43.45 (keep) and update_before has τ = −134.17 (archive); only the lineage flag separates them. The pair witnesses an opposite-sign collision; across all 225 worlds, the corresponding representation floors are 1.7196 and 0.0634. As an external check, the 78 LongMemEval knowledge-update questions give token-F1 0.4639 when the old session is archived, 0.2645 before the new session (−0.1994, CI [−0.2885, −0.1103]), and 0.3593 after (−0.1046, CI [−0.1891, −0.0216]). The old record contaminates the current answer, but the keep contexts are also longer and the order changes, so this is not a pure content intervention.

We do not claim cross-mechanism generalization. R3 varies geometry, not mechanism; LongMemEval is external but still an attribute-controlled synthetic corpus with no paired branch utility. The revision calls lineage a benchmark-specific diagnostic of version-conflict resolution, not evidence of transfer to independently designed mechanisms.

W2 / Q2 — Which claims refer to complete systems, and which to adapters

The displayed rows are matched-contract adapters. The abstract and §4.2 say this in the first sentence of the comparison, and the baseline audit (Table 11) separates native, ported, adapter and literature-only rows. We do not claim end-to-end superiority.

We add code-reproducible implementation checks on the full LoCoMo set (1,986 QA, shared Qwen3-8B reader). A-MEM (paper-faithful port) scores 0.2710, significantly above SQCAD auth-on (0.2425) and statistically indistinguishable from auth-off (0.2813). MemoryOS (port) scores 0.2266. Native Mem0 scores 0.0551 with Qwen and 0.0445 with DeepSeek. The full metric table is Table 12. A-MEM is not a native execution, so we do not claim QA SOTA. The result provides no evidence that the keep/archive decision carries downstream QA; we do not separately identify the contributions of view, packing and reader context. MemBase was not evaluated.

Q3 — "within 0.038" versus Terra 0.4294 vs 0.3139

We have rewritten the sentence to state the bound per backbone. On each backbone the better of the two reported guards is no more than 0.038 below the strongest adapter where the pair trails. The FIR guard alone trails by 0.0430 on Qwen, and on Terra SQCAD is ahead by 0.1155. The revised §4.2 states both the trailing bound and the per-guard exception; these main-table comparisons are descriptive, not significance claims.

Where to find the changes

  • Table 3: in-family decomposition; Table 16: leave-one-family-out refitting; Table 17: family-weight sensitivity.
  • Table 6 and Appendix B.2: authorization-removal ablation, including the note that the eligible set grows under auth-off.
  • Table 7 and Appendix B.3: certificate coverage, wrong-sign and the offline gated controller.
  • Table 12 and Appendix D: evidence tiers and implementation checks.

Thank you to the authors for the detailed rebuttal and additional experiments.

I appreciate the clarification regarding the distinction between the in-family representation information floor and the leave-one-mechanism-family-out regret. The additional sensitivity analysis and counterfactual examples help clarify the role of lineage information, while the explicit acknowledgment that these findings do not establish cross-mechanism generalization appropriately limits the scope of the contribution.

I also appreciate the more transparent characterization of the matched-contract baselines and the additional implementation checks involving A-MEM, MemoryOS, and Mem0. These results provide a more balanced perspective on the practical implications of the proposed framework. The clarification of the numerical discrepancy in Section 4.2 also addresses my third concern.

Although limitations remain, particularly regarding generalization beyond the constructed benchmark and end-to-end system-level benefits, I believe the rebuttal and revisions have meaningfully addressed my principal concerns.

I am therefore updating my rating from 2 to 3. My confidence remains unchanged.

Thank you for the careful reading and for updating your assessment. We are glad that the clarifications and additional analyses made the scope of the contribution easier to assess.

In the final version, we will preserve this framing: lineage information will be presented as a benchmark-specific diagnostic of version-conflict resolution, not as evidence of cross-mechanism generalization, and the baseline comparisons will remain explicitly limited to matched-contract implementations under a common protocol. We will also state explicitly the two limitations you note, namely generalization beyond the constructed benchmark and end-to-end system-level benefit.

Thank you again for the constructive engagement throughout the discussion period.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.