acceptodds
Under review as a conference paper at ICLR 2027

Memories Expire at Different Rates: Calibrating Agent Memory

Abstract

Agent memory can retrieve the right record, place it in time and revise a belief once a contradiction arrives. Yet none of these abilities tells whether a retrieved observation is still true before a contradiction appears. Recency heuristics discount age alone, but once memory types expire at different rates no function of age is calibrated for all types. We formalize the missing quantity as current validity, the probability that a stored proposition was correct when written and has not since been superseded, and build its estimator and decision rule together. Transition-Calibrated Memory (TCM) reads a memory’s transition context into a discrete-time invalidation hazard and integrates it over elapsed time into a probability. A cost-optimal rule then prices that probability into using or verifying the memory. We evaluate TCM on human re-query judgments, two intensive-care cohorts and an exact-event simulator. Conditioning on transition context is the robust gain: it cuts the worst group-calibration gap of age-only rules from 0.358 to 0.197 on TicToc, survives standard recalibration and largely disappears when the context is permuted. Beyond calibration, current validity also changes what an agent does. Routed through one cost-sensitive rule at a matched 20% verification budget, TCM earns 0.05 to 0.09 more harm-weighted utility than the strongest baseline on both clinical cohorts in every seed. By contrast, prompted and post-trained agents deciding alone verify out of proportion to risk, over-verifying eight probes in ten or acting on stale memories in three. Survival factorization is only one estimator of current validity, and its added value is regime-dependent, which separates the robust benefit of transition conditioning from the inductive bias used to estimate it.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
agent memory, current validity, confidence calibration, conditional hazard models, verification decisions

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_tPQmOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2(−1)
Read review
Paper summary

This paper studies how an agent should assess whether a stored memory remains valid before receiving contradictory evidence. It formalizes current validity as the probability that an observation was correct when recorded and has not subsequently been superseded. The proposed Transition-Calibrated Memory (TCM) estimates invalidation hazards conditioned on memory type, content, provenance, and available events, and combines them over elapsed time to obtain a validity probability. This estimate supports cost-sensitive decisions about whether to reuse or verify a memory. Experiments on TicToc, PhysioNet-2012, MIMIC-IV, and an exact-event simulator examine calibration, the contribution of transition context, the role of survival factorization, and downstream verification decisions. The results support conditioning on transition context over age-only heuristics, while showing that the additional benefit of survival factorization depends on the supervision regime. The paper also compares TCM with a direct validity predictor, TCM-D, to distinguish these two sources of improvement.

Critical strengths

The memory expiration issue in this paper is significant, which is important for applying this kind of memory system to daily assistant systems. And also I feel that the challenge proposed by the paper is suitable and insightful for solving this kind of problem: Two different memories with the same age may have different kinds of states because of their memory types and different objects. I also think that TCM solves this well; the solution is intuitive.

And I think the paper is well written and coherent, and also it's clear to read. I also like your experimental part. I think the experiments are solid. And in Section 5, they are well organized into different questions and provide enough findings. I think this is very interesting, and it covers the full process across different domains.

I like the problem formulation; it provides coherent ground truth for the paper.

Critical weaknesses

The relationship between the evaluation labels and the formal definition of current validity deserves further examination. TicToc provides human re-query judgments, which may reflect verification preferences as well as beliefs about whether an observation remains valid. The clinical datasets assess agreement at the next measurement, without observing all intervening changes. Although the paper acknowledges these limitations, they constrain how directly the calibration results can be interpreted as estimates of uninterrupted memory validity.

Questions for authors

What evidence supports interpreting TicToc re-query judgments as current-validity probabilities rather than preferences over verification actions?

For the clinical datasets, can the authors assess how departures from the original state followed by a return before the next measurement would affect the interpretation of the reported calibration results?

8 comments
AuthorsOpenReview ↗

Response to Reviewer tPQm

We appreciate your encouraging assessment of the problem formulation and experiments.

Consider glucose that is normal, becomes high, and returns to normal before the next measurement. The clinical label records endpoint agreement, although the original validity episode has ended. Agreement is the relevant target when the loss depends on the value at use. Uninterrupted survival is relevant when continuity itself matters.

What the human judgments support

The votes measure when people would reuse or re-check an observation. We now describe calibration there as calibration to the human reuse rate, one minus the re-query rate. Those votes combine judgments about validity with preferences about verification cost, so they do not identify a latent survival probability. TicToc tests the action-level prediction, while the simulator directly tests survival.

What the simulator lets us check

Exact simulator trajectories let us check whether the context result holds under either target interpretation.

The mapping after Definition 1 now makes this explicit:

DataObserved target
TicTocHuman reuse/re-query judgment
Clinical cohortsAgreement with the next measurement's band
SimulatorExact state survival and endpoint agreement

We quantified departures followed by returns in the simulator: 0.30% of original probes and 1.9% of observation-window probes in the seed-0 audit (Table 25). These frequencies are specific to that simulator and sampling design; intervening returns in the clinical cohorts remain unobserved.

The revised simulator also distinguishes seamless record renewals from actual state changes, using state survival S′ as its primary label and corrected interval-prefix event features. The comparison is:

Setting; Brier ↓Typed-ageTCM-DTCM
Trained and scored on state survival0.04540.01960.0236
Same predictions scored on endpoint agreement0.04800.03630.0382
Trained and scored on endpoint agreement0.04460.02800.0326

The ordering holds in every seed. Context improves prediction under both target interpretations, while the predicted probability has a distinct meaning in each (Figure 6b).

How this affects the clinical result

The clinical decision experiments address their observed endpoint target. At the same 20% verification budget, TCM improves utility by 0.052 and 0.093 on the two cohorts. The revision thus preserves the distinction you ask for while showing that the decision benefit and the simulator context result do not rely on treating these targets as interchangeable.

AuthorsOpenReview ↗

Follow-up on the revised target interpretation

Thank you again for your time and consideration of our work. We hope our response has helped connect the observed labels to the validity question that motivated the paper.

First, we now distinguish uninterrupted state survival from endpoint agreement explicitly. A value can change and later return, so agreement at the next measurement does not imply that it remained valid throughout. To check whether our finding depends on this distinction, we evaluated both targets in the simulator. Context improves prediction over typed-age under both targets in every seed. The advantage therefore does not depend on treating the two labels as interchangeable.

For the real data, we have revised the interpretation accordingly. TicToc supports calibration to observed human reuse judgments, while the clinical labels measure endpoint agreement and can miss a departure followed by a return. The simulator check does not recover those unobserved clinical transitions. This keeps each empirical conclusion tied to what its labels can establish.

We would be happy to discuss any remaining question about how these targets relate to the decision problem.

Thanks for the clear response and the added simulator analyses. Separating human reuse judgments, endpoint agreement and state survival makes the paper more precise, and I appreciate the honest reframing.

That reframing is also what changes my view. The quantity defined in Definition 1, uninterrupted validity, is now measured only in the simulator. TicToc measures human reuse rates, and the clinical cohorts measure agreement at the next observation. So the real-data calibration results no longer speak directly to the paper's core estimand.

In the simulator, where that estimand is observed, the results are less favorable to TCM than I expected:

TCM-D beats TCM in every matched setting and every seed. TCM's worst-group Brier is clearly worse than the typed baselines: 0.305 against 0.171 for typed-age in Table 3, and 0.408 against 0.229 for typed constant decay with exact episodes. The paper is motivated by calibration across memory types, so this matters. The original event features used episode-end information, which an early interval should not see. That was only fixed during the rebuttal, so I am less sure how far the earlier simulator results can be trusted. On the real data, TCM ties the flexible typed-age hazard on TicToc (0.1368 overall), and its worst-group Brier is slightly worse on both TicToc and MIMIC-IV. The clinical decision gains are real, but they target endpoint agreement rather than the validity notion the paper sets out to estimate.

I still think the problem is important and context conditioning is a useful finding. However, the main estimator and the paper's central estimand are less well supported after the revision than I had assumed in my review. I am lowering my score to 2.

Reviewer_9rrsOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies the question of whether a stored memory remains valid when retrieved, distinguishing current validity from relevance and original correctness. The authors propose Transition-Calibrated Memory (TCM), which estimates context-dependent invalidation hazards and combines them with origin reliability to guide cost-sensitive use, verification, or abstention; They also propose TCM-D that predicts validity directly. Controlled experiments on clinical datasets, TicToc, and simulated transitions support context-conditioned prediction over age-only baselines and show utility gains at matched verification budgets. The main contribution is the problem formulation and empirical investigation, with survival factorization helping primarily in particular supervision regimes.

Critical strengths
  1. The studied question is meaningful and timely
  2. Strong experimental controls. The authors include flexible age baselines, direct prediction, recalibration, and context-permutation tests, while acknowledging where survival factorization adds little.
  3. Improved decision utility. The results indeed show that at a 20% verification budget, TCM improves harm-weighted utility by approximately 0.052 on PhysioNet-2012 and 0.093 on MIMIC-IV, enabling better decisions without additional verification.
Critical weaknesses
  1. Labels do not directly measure memory validity. The formal target requires that no superseding transition has occurred, but clinical labels measure agreement with the next recorded measurement, and TicToc uses judgments about re-querying. Endpoint agreement can hold after a normal -> abnormal -> normal trajectory even though the original episode expired. These results therefore concern more about observed agreement or re-query judgments, which differ from the survival probability the authors used in the decision framework.
  2. Incomplete censoring comparison. The claim that direct prediction cannot use censored episodes seems a bit too strong for me: a correctly initialized episode censored at time c provides a positive survival-status observation at c. TCM and TCM-D perform almost identically when trained on the same labeled data (Table 16: 0.0241 versus 0.0239). Hence, TCM’s better performance in the censoring experiment may come from using additional supervision from censored episodes instead of survival factorization itself.
  3. Limited robustness to distribution shifts. On unseen TicToc transition families, TCM’s worst-group Brier score rises to approximately 0.412, versus 0.261 for the flexible age-only hazard. The post-hoc fallback detects explicit unseen types or out-of-range elapsed times and activates on every probe in the tested shifts which does not directly address changing transition rates within familiar types and observed time ranges. The authors are encouraged to explain more here.
  4. Unclear evidence of improved calibration. Lower Brier score can reflect better discrimination as well as calibration. On PhysioNet-2012, TCM’s maximum group-calibration gap is 0.158, versus 0.096 for the flexible typed-age hazard, despite its better overall Brier score. The authors are encouraged to separate these effects which would clarify the claimed calibration gains.
Questions for authors
  1. How do the conclusions change when exact simulator transitions are used to compare episode-survival and endpoint-agreement labels at independently sampled probe times?
  2. Does TCM retain its censoring advantage when TCM-D receives the same usable pre-censor validity supervision and training budget?
  3. How do TCM, TCM-D, and the fixed fallback perform when transition rates change within familiar types and observed time ranges?
3 comments
AuthorsOpenReview ↗

Response to Reviewer 9rrs (1/2)

Thank you for recognizing the decision benefit and for the concrete suggestions on labels, censoring, rate changes and calibration.

Both contextual models outperform typed-age on both target labels and under matched supervision. TCM-D has the lower Brier in the matched comparison.

What changes when the label changes (W1 / Q1)

The revised mapping after Definition 1 distinguishes TicToc's human reuse judgments and the clinical endpoint labels from uninterrupted state survival. Exact simulator trajectories let us test both survival S′ and endpoint agreement E.

A departure followed by a return has S′=0, E=1. In the seed-0 audit, this occurs in 0.30% of original probes and 1.9% of observation-window probes. The corrected S′ label also keeps a state valid through seamless medication renewals. Probe windows no longer depend on the target record's stop time, and each hazard interval uses only events visible by its start. The patient observation window still includes death. Appendix R documents these changes and the submitted target for reproduction.

To test whether the context finding depends on the target definition, we compared three settings. First, after refitting on S′, Brier is 0.0236 (TCM), 0.0196 (TCM-D) and 0.0454 (typed-age). Second, scoring those same predictions on E gives 0.0382, 0.0363 and 0.0480, isolating the change in evaluation target. Finally, training directly on E gives 0.0326, 0.0280 and 0.0446. The ordering holds in every seed (Figure 6b), so the context advantage survives both target interpretations, including when the model is trained for endpoint agreement.

What changes when supervision is matched (W2 / Q2)

A separate question is whether unequal supervision explains the comparison between TCM and TCM-D. TCM-D already used pre-censoring probes, and we have corrected the statement that it could not consume censored episodes. To test the remaining supervision difference directly, we added a shared-grid control: both models receive the same usable state labels, equal weight per record and the same probe-time loss.

Matched settingTyped-ageTCM-DTCM
Probe labels: Brier0.04540.02000.0239
Probe + grid labels: Brier0.04550.02070.0247
Probe + grid labels: worst-group Brier0.1740.1710.259

TCM-D has lower overall Brier than TCM in every seed, with paired patient-bootstrap intervals excluding zero in all five. Supplying exact episodes through TCM's native likelihood changes its Brier little, from 0.0239 to 0.0236. TCM is also worse on the worst-group metric above.

What the matched control establishes

The matched comparison identifies context as the source of the common improvement. TCM-D was included in the original design precisely to separate context from survival parameterization. Table 3 and Section 5.4 now state the result of that comparison, and the factorization claim has been revised accordingly.

AuthorsOpenReview ↗

Response to Reviewer 9rrs (2/2)

Changing transition rates within familiar support (W3 / Q3)

We rescaled durations of a seeded half of named record types by 2 or 0.5. Start times, probes, observation window, trained models and fallback rule stayed fixed, so the intervention changes the test process. Five-seed mean Brier is reported in Table 26:

Duration scaleTyped-ageTCM-DTCM
Original0.04540.01960.0236
×20.04760.03550.0338
×0.50.15170.13470.1349

Both contextual models beat typed-age in every seed under both interventions. At doubled durations, TCM also beats TCM-D in every seed, with the paired interval excluding zero in four of five. At halved durations, all models degrade substantially, and the two contextual predictors have similar Brier. Their 20%-budget utilities are about 0.56, versus 0.51 for typed-age.

Fallback fires on only 0.68% of probes in both interventions. This answers your distinction: it detects lack of support, rather than rate drift within familiar support. The revision states that boundary explicitly. The doubled-duration result indicates a regime-dependent benefit from factorization, while the opposite intervention does not show that benefit.

Separating calibration from discrimination (W4)

Turning to calibration, lower Brier alone does not identify whether reliability or resolution improved. We therefore decomposed the saved clinical predictions into ten bins. Relative to typed-age, the results attribute the gain to resolution (Table 24):

Data / methodBrier ↓Reliability ↓Resolution ↑
PhysioNet-2012: TCM0.14770.000160.0168
Typed-age0.15880.000040.0054
MIMIC-IV: TCM0.12330.000180.0248
Typed-age0.14150.000160.0067

The resolution difference is supported in 5/5 PhysioNet and 15/15 MIMIC seeds. Reliability is not better than typed-age, and the group-gap comparison you identified remains: 0.158 for TCM versus 0.096 for typed-age on PhysioNet. TCM and typed-age do not separate in the TicToc decomposition.

Relative to age-only, TCM improves resolution on all three datasets. It also improves reliability on PhysioNet and reduces the maximum group gap on TicToc and MIMIC, each in every seed. Section 5.2 now states the comparator and metric for each conclusion.

The clinical utility gains remain 0.052 and 0.093 at the same 20% verification budget. The decomposition explains how context supports that decision benefit without equating lower Brier with uniformly better calibration.

AuthorsOpenReview ↗

Follow-up on the revised analyses and interpretation

We appreciate the time and care you have put into reviewing the paper. We hope our response has helped connect the additional analyses to your questions about supervision, calibration and distribution change.

  1. First, to separate context from supervision and factorization, we compared the contextual models under matched supervision and tested both simulator targets. Both beat typed-age, while TCM-D has lower Brier than TCM in the matched setting. We therefore corrected the censored-data statement and revised the factorization interpretation.
  2. Next, to explain the lower clinical Brier, we decomposed it into reliability and resolution. The advantage over typed-age comes from better resolution, not better reliability. We accordingly revised the calibration wording so it reflects what the decomposition actually establishes.
  3. Finally, to examine changes within familiar support, we tested both duration interventions. Both contextual models beat typed-age in every seed under both interventions. With doubled durations, TCM also has lower Brier than TCM-D in all five seeds (means 0.0338 versus 0.0355), with paired intervals excluding zero in four of five seeds. With halved durations, the two contextual models perform similarly and all methods deteriorate substantially. This also clarifies why support fallback need not activate: it detects unfamiliar support, not a rate change within it.

At the same 20% verification budget, TCM improves clinical utility by 0.052 and 0.093. The additional analyses show that the context advantage persists under matched supervision, both target definitions and both tested duration shifts. We would be happy to discuss any remaining questions about these comparisons.

Reviewer_uY59OpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies whether a stored memory remains valid as time passes. Transition-Calibrated Memory (TCM) uses memory type, text, provenance, and available events to estimate an invalidation hazard. It converts this into a validity probability and uses a cost-based rule to choose whether to reuse or verify the memory. Experiments use human re-query judgments, two clinical datasets, and a simulator. Context improves prediction and verification decisions. The extra benefit of the survival model depends on the training data and supervision.

Critical strengths
  1. The paper addresses a useful memory problem. A relevant record can still be outdated. Modeling validity by memory type and context gives agents a practical way to decide which records need checking.

  2. The gains are supported by careful controls. The paper tests recalibration, context permutation, and a direct predictor. At a matched 20% verification budget, TCM improves utility by 0.052 and 0.093 on the two clinical datasets (Section 5.5).

Critical weaknesses
  1. The real-data labels do not directly measure episode survival. TicToc labels reflect whether people would re-query. Clinical labels compare values at two measurements and can miss changes followed by a return. These labels do not establish that a memory was never superseded, as Definition 1 requires.

  2. The survival-model comparison mixes several factors. TCM and TCM-D nearly tie on real data. Some simulator gains use censored episodes excluded from TCM-D, while other survival baselines use different training objectives. The comparison needs tighter controls to isolate the benefit of survival factorization.

Questions for authors
  1. How sensitive are the conclusions to the difference between episode survival and validity at the probe time? On the simulator, could you score both definitions, including cases where the state changes and later returns? This would help show how well the real-data labels support the formal target.

  2. Why can TCM-D not use censored episodes? A direct time-conditioned predictor could receive valid labels at times before censoring, including the censoring time. Could you compare this version with TCM using the same episodes and comparable probe-time loss weights? The survival baselines should also be selected for the same probe-time objective.

2 comments
AuthorsOpenReview ↗

Response to Reviewer uY59

Thank you for asking for the information-matched comparison. Both contextual predictors retain their advantage over age rules, while TCM-D has the lower Brier under matched supervision. The new label analysis also shows that the context result holds for both state survival and endpoint agreement.

On the target labels (your first question)

The text after Definition 1 now distinguishes human reuse judgments on TicToc, endpoint agreement on the clinical cohorts, and exact state survival on the simulator. A clinical endpoint label leaves intervening transitions unobserved.

In the simulator we can compare record survival S, state survival S′ and endpoint agreement E directly. A normal → abnormal → normal trajectory has S′=0 and E=1. Such returns occur in 0.30% of the original probes and 1.9% of observation-window probes in the seed-0 audit (Table 25). These frequencies describe the simulator, rather than unobserved returns in the clinical cohorts.

The audit also separated seamless medication renewals from state changes. The corrected S′ target keeps a state valid through a seamless renewal. New probe windows avoid dependence on that record's stop time, and hazard intervals use events visible by their start. Appendix R records the corrections and retains S for reproduction.

To check whether the context advantage survives these target and feature corrections, we refitted on S′. Brier is 0.0236 for TCM, 0.0196 for TCM-D and 0.0454 for typed-age. We then scored the same predictions on E, obtaining 0.0382, 0.0363 and 0.0480. The ordering holds in every seed. Training directly on E also preserves the context advantage, so the finding is not confined to scoring predictions against the state-survival target (Figure 6b).

On giving the two models the same information (your second question)

You are right that censored episodes provide valid pre-censoring labels. TCM-D already used such probes, and we have corrected “cannot consume censored episodes.” The new comparison supplies both models with the same state labels at frozen grid edges, equal total weight per record and the same probe-time loss. No labels are assigned beyond censoring.

Supervision; five-seed BrierTyped-ageTCM-DTCM
Same probe labels0.04540.02000.0239
Same probe + grid labels0.04550.02070.0247

TCM-D's lead over TCM is about 0.004 in both settings, with paired intervals excluding zero in all five seeds. Adding exact episodes to TCM's native likelihood changes its Brier only from 0.0239 to 0.0236. These controls attribute the common improvement to context, rather than an additional benefit from survival factorization in these settings (Table 3).

What this means for the decision result

The decision benefit remains: corrected exact-episode utilities at a 20% verification budget are 0.895 for TCM and 0.901 for TCM-D, versus 0.848 for typed-age. This complements the unchanged clinical gains you highlighted.

For the survival baselines, the Cox penalty selected by development partial likelihood also minimizes probe Brier in the submitted grid (Table 13). DeepSurv and DeepHit retain their library configurations and own early-stopping losses.

AuthorsOpenReview ↗

Follow-up on the matched-supervision revision

Thank you again for your time and thoughtful review. We hope the matched-supervision comparison has helped clarify where the improvement comes from.

First, we corrected the statement about censored data: TCM-D already used valid pre-censoring probes. To remove unequal supervision as an explanation, we then compared both models using identical usable labels and per-record weights in the corrected simulator's shared-grid setting. Brier is 0.0207 for TCM-D and 0.0247 for TCM, compared with 0.0455 for typed-age. Both contextual predictors therefore retain their advantage over the age rule. Since TCM-D is better in this comparison, we attribute the shared improvement to context rather than an added factorization advantage in this setting.

We also checked whether that finding depends on the target definition by evaluating state survival and endpoint agreement separately. The context advantage remains under both targets, supporting the conclusion beyond one label interpretation.

We would welcome your thoughts on the matched comparison and would be happy to clarify any remaining detail.

Comments on the paper

3 comments
AuthorsOpenReview ↗

General response: decision utility and matched-context comparisons

Thank you for the thoughtful reviews. The additional comparisons clarify both the decision benefit and its source.

The decision benefit

At the same 20% verification budget, TCM improves mean utility by 0.052 on PhysioNet-2012 and 0.093 on MIMIC-IV, with improvements in every seed. Recalibration and context-permutation controls support the role of transition context (Sections 5.3 and 5.5).

What the added controls tell us

We used the corrected simulator and matched-information tests to separate the contribution of transition context from the choice of estimator. With identical shared-grid labels, both contextual models outperform typed-age: Brier is 0.0207 for TCM-D, 0.0247 for TCM and 0.0455 for typed-age. TCM-D nevertheless has lower Brier than TCM in every seed. These controls identify context as the source of their common improvement in this setting, and we have revised the factorization claim accordingly.

Clarifications in the revised paper

We have also made the observed targets explicit after Definition 1: human reuse judgments, agreement at the next clinical observation, and simulator state survival are distinguished. Figure 6 compares alternative labels and duration interventions. Appendix R records the simulator corrections and full analyses; the individual replies address label interpretation, censoring, rate changes and calibration. The revised PDF accompanies these replies. We have added blue markings to make key revisions in the main text and appendix easier to find.

AuthorsOpenReview ↗

Response Summary: Validity Targets, Matched Controls and Decision Utility

Dear Reviewers,

Thank you for the thoughtful questions and for recognizing the usefulness of context-dependent memory validity. The original experiments showed that this information improves decisions: at the same 20% verification budget, TCM increases harm-weighted utility by 0.052 on PhysioNet-2012 and 0.093 on MIMIC-IV. These results remain in the paper. Our responses address how the labels and model comparisons connect to this finding.

Which notion of validity does each dataset observe?

For tPQm, uY59 and 9rrs, we distinguish uninterrupted episode validity, agreement at the query time, and human re-query judgments. A normal–abnormal–normal trajectory illustrates why the first two differ. The simulator checks both targets directly. This ties each empirical result to its observed target and the decision loss it can inform.

Does the context benefit survive matched supervision?

To answer uY59 and 9rrs, we compared TCM and the direct predictor TCM-D using the same labels, inputs and probe-time loss weights. With the corrected event features and state-survival labels, their Brier scores are 0.0247 and 0.0207, respectively, versus 0.0455 for the matched typed-age control. Both context-conditioned estimators retain the advantage over age-based prediction. The comparison identifies context as the shared source of the gain and tests the additional role of survival factorization separately.

What happens when durations change?

For 9rrs, we tested duration changes within familiar support, where a support detector should not be expected to signal drift. Both context-conditioned estimators outperform the typed-age control in Brier and decision utility in each tested setting and seed. When durations double, TCM also has lower Brier than TCM-D in all five seeds.

Together with the original recalibration and context-permutation controls, these checks explain where the memory-validity gains come from. We would be happy to discuss any remaining question about the targets or their use in verification decisions.

AuthorsOpenReview ↗

Summary of Major Contributions, Strengths, and Responses to Concerns

Dear AC, SAC, and PCs,

We sincerely appreciate your time and effort in managing the review process. We also thank the reviewers for their thoughtful assessments, including the positive recommendations from uY59 and 9rrs, and tPQm’s recognition of the importance of memory expiration and the usefulness of context conditioning. To assist your final assessment, we summarize the paper’s contributions, the strengths recognized in the reviews, and the evidence addressing the main concerns.

1. Core Contributions

  • A formulation of current validity for agent memory. Retrieving a relevant record does not establish whether it remains valid before contradictory evidence arrives. We formalize this missing quantity and model how validity changes with a memory’s age, type and transition context.

  • An estimator connected directly to verification decisions. TCM converts context-dependent invalidation hazards into validity probabilities, which a cost-sensitive rule uses to choose between reusing and verifying a memory. At the same 20% verification budget, TCM improves harm-weighted utility over the strongest baseline by 0.052 on PhysioNet-2012 and 0.093 on MIMIC-IV, with improvements in every seed.

  • Controlled evidence that transition context matters beyond recency. Flexible age baselines, recalibration and context-permutation controls distinguish the contribution of context from model capacity and score adjustment. On TicToc, TCM reduces the maximum group-calibration gap from 0.358 to 0.197 relative to the flexible age-only hazard. The direct predictor TCM-D further separates this context benefit from the choice of survival parameterization.

2. Strengths Recognized by Reviewers

  • A meaningful and practically relevant problem. uY59 highlighted that a relevant memory can still be outdated. tPQm recognized the importance of memory expiration for assistant systems and the intuition that memories of the same age can require different treatment. 9rrs described the question as meaningful and timely.

  • Careful controls and useful decision gains. uY59 and 9rrs highlighted the recalibration, context-permutation and direct-prediction controls, together with the matched-budget utility results. tPQm also appreciated the coherent presentation and experiments across domains.

3. Responses to Key Concerns

The additional analyses test whether the original context and decision findings survive changes in the target, supervision and transition process.

ReviewerConcernEvidence in the final manuscript
tPQm, uY59, 9rrsHow do the observed labels relate to current validity?The paper distinguishes human reuse judgments, clinical endpoint agreement and exact simulator state survival. Training and evaluating separately on state survival and endpoint agreement preserves the context advantage in every seed (Figure 6).
uY59, 9rrsDoes unequal supervision explain the gains?With identical shared-grid labels and per-record weights, both contextual predictors outperform typed-age in every seed. This isolates the benefit of context from an additional benefit of survival factorization (Table 3).
9rrsWhat happens when transition rates change within familiar support?Under both tested duration interventions, both contextual predictors outperform typed-age in overall Brier and decision utility in every seed. When durations double, TCM also has lower Brier than TCM-D in all five seeds (Table 26).

4. Concluding Remarks

The paper connects a missing quantity in agent memory to a practical decision: how much to trust a stored observation, and when verification is worth its cost. The original controls and the additional matched experiments support transition context as a useful source of information, while the clinical results demonstrate better decisions without additional verification. We appreciate your careful consideration of these contributions.

The authors of TCM

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.