acceptodds
Under review as a conference paper at ICLR 2027

Learning What to Remember: Utility-Guided Memory for Long-Horizon Text-to-SQL

Abstract

Memory systems have proven beneficial in a variety of long-horizon settings. As text-to-SQL research increasingly extends to multi-turn interactions, these systems must retain user-specific knowledge across turns, creating a need for effective memory management. Text-to-SQL presents both opportunities and challenges for memory: each completed round yields a structured record comprising a request, SQL program, execution outcome, and possible correction, while interactions tend to revisit a bounded set of schema-grounded concepts. However, data reflecting real user workloads in long-horizon, multi-turn settings remains limited. Leveraging structured interactions, we represent completed rounds as standardized memory entries and formulate bounded text-to-SQL memory management as an eviction problem. We further exploit schema concentration through P-UTL, a lightweight, text-to-SQL-native policy that predicts the utility of stored information using execution-based supervision. To address the lack of workload simulators, we introduce a controlled trajectory generator that varies dormancy, recurrence, information density, and memory pressure across extended SParC- and CoSQL-derived streams. Experiments across workload configurations and baselines demonstrate the advantages of our method and show that predictive utility is particularly valuable when useful information competes for bounded memory.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Text-to-SQL, Memory Management, Long-Horizon Agents

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_EGyUOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes P-UTL, a memory management method for long-horizon Text-to-SQL. The core idea is to maintain a fixed-size memory pool of entries, and use a lightweight MLP to predict each entry's future utility. When the pool is full, only the top-B entries are retained. The MLP is trained offline on SQL execution data: removing an entry and re-running the query to see if accuracy changes. Experiments on synthetic scenarios show P-UTL outperforms baselines including LRU and LLM-based manager, especially under high memory pressure.

Critical strengths

S1. Long-horizon conversation memory is an important and underexplored problem.​ Real-world multi-turn Text-to-SQL systems frequently encounter situations where early context (e.g., table definitions, business rules) must be recalled dozens of turns later. Existing methods either keep everything (costly) or use generic heuristics like recency. P-UTL addresses this gap with a principled approach: learning which entries actually matter based on downstream task performance. Even if the proposed method has limitations, the problem formulation itself advances the field.

Critical weaknesses

W1. Single-pass eviction has an irreversible information loss problem with no mitigation.​ Once an entry is evicted, it is permanently gone and cannot be retrieved even if later queries depend on it. This is an architectural flaw and fundamentally limits the method's applicability to truly long conversations where early context may become relevant again after many turns.

W2. The validity of the evaluation dataset is questionable.​ The entire benchmark is synthetically constructed: recall schedules follow hand-picked distributions (Normal-35, Poisson-40). There is no evidence that these distributions reflect real user behavior in long-horizon Text-to-SQL conversations. The paper does not report how conventions are phrased (template vs. free-form). Without validation on real or more diverse data, the reported improvements may be artifacts of the synthetic setup rather than genuine advances in memory management.

Questions for authors

Q1.​ Have you tested P-UTL in the adversarial setting where new concepts are continuously introduced, their total number far exceeds the memory budget B, and queries uniformly sample from all historical concepts? If so, what were the results? If not, can you provide a theoretical analysis of how P-UTL would behave in this case?

Q2.​ How were the conventions (definitions) generated? If they follow a template (e.g., "define X as Y"), does the MLP simply learn template-specific cues? Do you have any results on free-form, human-written definitions, or cross-database generalization?

Q3.​ Are the training and test datasets generated from the same synthetic pipeline? If so, could the MLP be overfitting to distributional regularities of that pipeline rather than learning a general notion of utility? Have you validated P-UTL on data from a different generation process, such as different recall distributions or real human conversations? Without such cross-distribution validation, how can we rule out that P-UTL's advantage over LRU is largely an artifact of overfitting to the specific synthetic distribution used in both training and evaluation?

2 comments
AuthorsOpenReview ↗

Response to Reviewer EGyU

We thank the reviewer for their thoughtful feedback. Below, we clarify the convention-generation procedure and present additional experiments addressing irreversible eviction and generalization beyond fixed definition templates. We welcome the reviewer’s further feedback on these responses.

1. Irreversible eviction under continuous concept arrivals

The reviewer is correct that an evicted entry cannot be recovered from the bounded store, and we do not claim that P-UTL can recall every historical concept indefinitely. When future requests may concern any past concept, a finite memory pool cannot guarantee perfect recall. The fair comparison here is how different policies perform under the same memory constraint.

To evaluate the proposed adversarial setting, we construct 100-round trajectories in which new definitions and their associated warm-up turns are introduced densely throughout the interaction (see Figure 3 for the definitions of these turn types). We randomly insert 10 callbacks between turns 26 and 100, with each callback referring to a definition randomly selected from all preceding definitions. We generate three such trajectories for evaluation on the department_store database.

[begin-of-traj][concept][warm-up][dormancy]...[concept][warm-up][dormancy]...[concept][callback(randomly inserted)][warm-up]...[callback]...[callback][end-of-traj]
P-UTLE-LLMLRU
Continuous concept arrivals76.6746.6743.33

P-UTL outperforms E-LLM and LRU even when the number of definitions exceeds the memory budget and callbacks sample the full available history. These results demonstrate an advantage beyond the original recurrence schedule.

2. How the submitted conventions are generated

The convention workloads in the submitted paper use a fixed template. Each definition specifies that a named term denotes rows satisfying a numeric predicate and includes an initial COUNT query. This is followed by three warm-up queries, AVG, COUNT(DISTINCT), and MAX, in shuffled order, and a delayed SUM callback. An example is shown below:

[Definition] For the rest of our conversation, a “high-value order” is a row in Orders where order_total > 1000. Use this definition. To start: how many high-value order entries are there? 
[Reference SQL] SELECT COUNT(*) FROM Orders WHERE order_total > 1000;
[Warm-up Questions]
What’s the average order_total among the high-value orders?
How many distinct values of order_total appear in the high-value orders?
What’s the maximum order_total we’ve seen among the high-value orders?
[Callback] Going back to our earlier definition: what is the SUM of order_total across all high-value orders? 

To examine generalization beyond this template, we conduct an additional experiment using free-form definitions, as described in the next section.

3. Generalization to free-form definitions

To test whether P-UTL’s performance depends on the fixed definition wording, we replace the templated definitions with LLM-generated free-form descriptions of the same executable predicates and check their semantic fidelity. We keep the initial questions, warm-up queries, callback operations, and recurrence patterns unchanged. An example is:

A premium price refers to the cost of a product listed in the Products table that exceeds 815, suggesting a higher value or quality compared to other items. To start: how many premium price entries are there? 

We regenerate the workload using Poisson distribution (λ=40\lambda=40), normal spacing, and 30 callbacks. We evaluate the pretrained P-UTL-FULL predictor directly, without retraining.

P-UTL-FULLE-LLMLRU
Free-form workload78.964.431.1

P-UTL-FULL handles free-form definitions without retraining, outperforming E-LLM and LRU by a large margin. Since these definition forms are absent from its training data, the results provide evidence that P-UTL generalizes beyond the specific definition template used in training.

Dear Reviewer,

While the rebuttal is ending soon, we looking forward to your response to check whether our rebuttal addressed your concerns. We would be happy to make clarifications if you had concerns unresolved.

Best,

Submission 22100 Authors

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

The paper studies memory admission and eviction for long multi-turn text-to-SQL. Each completed round becomes a memory entry. In the fixed-slot comparisons, methods share the SQL answerer, top-five retriever and memory writer, and differ in which entries they retain.

P-UTL is a small MLP trained from offline counterfactual SQL executions. Removing a retrieved memory reveals whether it helps, harms or leaves later execution correctness unchanged. These effects are accumulated into training labels. At inference, the predictor uses fixed write-time features to score entries and retains the highest-scoring ones.

The evaluation stitches SParC/CoSQL dialogues into streams of about 200–650 turns. Verbatim callbacks repeat earlier questions after uniformly sampled gaps. Convention callbacks reuse user-defined numeric predicates after exponential, normal or Poisson gaps.

Critical strengths
  1. Counterfactual SQL execution gives memory retention a task-level training signal. Under capacity and density pressure, P-UTL achieves 94.2% accuracy versus 65.9% for E-LLM (Table 3). Its online memory updates take about 2 ms without manager API calls (Table 5), making the improvement practical rather than dependent on a costly manager.

  2. The evaluation varies dormancy, definition density and memory capacity on shared trajectories. Convention callbacks require applying an earlier predicate to a different aggregation (Listing 2). Mixed-workload training retains comparable accuracy on held-out databases (Table 4), and Tables 14–15 check that callback gains do not come at the expense of ordinary queries relative to the stateless control.

Critical weaknesses
  1. The experiments do not sufficiently isolate the contribution of execution-based supervision. Training and testing share the definition, warm-up and SUM-query pattern (Appendices A.1 and A.3). The episodic-priority heuristics in Table 9 perform poorly, but do not directly test a transcript-only rule that specifically prioritizes explicit definitions. Without this more targeted control, the gains cannot be clearly attributed to learned utility rather than recognition of definition rounds. The cross-database results are useful, but transfer beyond this recipe remains untested.

  2. The training labels measure removal from a retrieved prompt, not the consequences of eviction from a bounded pool (Appendix A.2). The replacement-aware diagnostic in Appendix C.3 is helpful, but 101 of 305 originally nonzero labels become neutral in its stratified sample (Table 17). No policy-performance comparison using replacement-aware training labels is reported. Preserving label polarity therefore does not establish that retention rankings or downstream accuracy are unaffected; describing the change as merely diluting supervision is stronger than the evidence supports.

  3. The Scenario-2 advantage over E-LLM, 95.9% versus 92.8%, lacks a paired uncertainty estimate. Figure 2 reports trajectory-bootstrap intervals, but not for this headline difference. An unbounded store with the same top-five retriever is also missing as a reference. These omissions make the smaller gains and the effect of capacity harder to assess, although they do not negate the large Scenario-3 improvement.

Questions for authors
  1. How does P-UTL compare with a rule that prioritizes explicitly introduced definitions using only transcript information, without generator metadata?
  2. Does training on replacement-aware labels change retention decisions or callback accuracy? If this comparison is unavailable, please qualify the conclusion drawn from Table 17.
  3. Please report unbounded-store accuracy with the same top-five retriever and a paired trajectory-bootstrap interval for the Scenario-2 difference between P-UTL and E-LLM.
  4. Appendix C.1 notes possible non-monotonicity. For Poisson-40 with dense spacing and 15 conventions, accuracy falls from 99.4% at 16 slots to 80.0% at 32 (Tables 3b and 13b). Is the same predictor used, and what explains this larger change?
  5. Please quantify offline labeling and training costs, which are acknowledged in Section 5 but excluded from Table 5.
3 comments
AuthorsOpenReview ↗

Response to Reviewer CQHT (1)

We thank the reviewer for their constructive comments and insightful feedback. We have conducted additional experiments to address the concerns about definition-aware baselines, replacement-aware supervision, retrieval, and offline computational cost.

1. Explicit definition-concentrated baseline

Our E-LLM baseline already incorporates definition-aware instructions. Specifically, its memory-management prompt explicitly prioritizes conventions, definitions, business terms, and thresholds:

You are a memory manager for a text-to-SQL agent. You will be shown a numbered list of memories currently held in the agent's memory pool. Your job: pick the memories that are most useful to keep for answering future SQL questions. Keep memories that state non-derivable facts (conventions, definitions, business terms, thresholds) or that successfully answered a similar question. Discard duplicates and routine noise. Directly provide the final answer without analysis or explanation. Respond ONLY with a JSON array of the 1-based indices you choose to KEEP; prioritize emitting the JSON answer immediately.

To provide a more targeted control, we additionally evaluate Def-priority, a transcript-based heuristic that extracts explicit definitions, deduplicates their predicates, and prioritizes their retention. As detailed in our response to Reviewer WnzZ, Def-priority matches P-UTL-FULL under the original convention form, but when lower-bound conventions are replaced with unseen interval predicates and the callback aggregation is changed, Def-priority drops to 71.7, whereas P-UTL-FULL achieves 98.3. Results on the more difficult setups further support the advantage of our methods. We consider this exploration shows our method’s capability of generalization beyond the original convention pattern. Appendix C.1 and Table 9 provide complementary comparisons with broader episodic-memory retention heuristics.

2. Replacement-aware supervision and downstream performance

To directly assess whether the label changes identified in Appendix C.3 affect downstream performance, we retrain P-UTL on the full training set using replacement-aware labels. The results are:

ScenarioOriginal-label P-UTLReplacement-aware P-UTL
Scenario 295.9196.10
Scenario 394.2492.77

Replacement-aware supervision changes performance very slightly in both scenarios. These results provide direct evidence that the observed label changes do not substantially affect downstream accuracy in the tested settings.

3. Unbounded-store reference and paired uncertainty

Following the reviewer’s suggestion, we evaluate an unbounded memory store using the same top-five retriever at query time.

P-UTLE-LLMUnbounded store
Scenario 295.9192.7598.33
Scenario 394.2465.9168.90

The unbounded store exceeds P-UTL by 2.42 points in Scenario 2, but falls below it by a large margin in Scenario 3. For the former scenario the results are plausible, since the unbounded store prevents the memory being permanently lost during evictions. For the latter, We hypothesize that the larger set of competing memories in Scenario 3 makes it harder for the retriever to select the relevant entries. To examine this possibility, we report the unbounded baseline’s accuracy by callback count and trajectory configuration:

# of callbacksNormal μ=35\mu=35, normalNormal μ=35\mu=35, densePoisson λ=40\lambda=40, normalPoisson λ=40\lambda=40, dense
1581.2581.2574.3872.50
2073.5075.5067.0066.50
2568.6467.7366.3665.45
3066.6764.1762.5060.83

Accuracy decreases as the configured callback count increases across all four configurations. This pattern is consistent with the retrieval-competition hypothesis.

Bootstrapping. For Scenario 2, the paired bootstrap 95% confidence interval for the accuracy difference between P-UTL and E-LLM is **

−0.52,+6.85−0.52, +6.85

** percentage points. We will report this uncertainty alongside the point estimates.

AuthorsOpenReview ↗

Response to Reviewer CQHT (2)

4. Trajectory matching and the split correction

We investigated the non-monotonic performance when increasing the memory budget from (B=16) to (B=32), comparing the retained and retrieved memories at matched rounds.

  1. The (B=16) memory pool was a subset of the (B=32) pool at every examined round. Thus, the larger pool retained all entries available in the smaller pool.
  2. However, the top-five retrieved sets shared only 2.01 memories on average, indicating substantial differences in which retained entries reached the SQL answerer.
  3. For queries answered correctly at (B=16) but incorrectly at (B=32), inspection showed that the smaller pool supplied the critical memory through retrieval, whereas the larger pool did not.

These observations indicate that the performance drop arises from retrieval differences: retaining the critical memory does not guarantee its inclusion in the final top-five context. Potential remedies include improving the retriever, increasing the retrieval budget, or providing the full memory pool when the answerer’s context budget permits. While our primary contribution concerns memory management instead of retrieval, we will add this analysis to the manuscript and clarify that end-to-end performance depends on the interaction between memory retention and retrieval methods.

5. Offline SQL calls and computational cost

Dataset construction required 507,973 calls for the training set and 223,356 calls for the validation set, totaling 731,329 API calls. Each call used an average of approximately 627 input tokens and 22 output tokens, with aggregate usage of approximately 474.91 million tokens. Using GPT-4o-mini, the estimated API cost of constructing the dataset was USD 78.45. We plan to release the trajectories after the review process.

The P-UTL predictor is a three-layer MLP with 245,507 trainable parameters and takes less than one hour to train on a single NVIDIA A6000 GPU. These offline annotation and training costs are separate from the online memory-management costs reported in Table 5.

Dear Reviewer,

While the rebuttal is ending soon, we looking forward to your response to check whether our rebuttal addressed your concerns. We would be happy to make clarifications if you had concerns unresolved.

Best,

Submission 22100 Authors

Reviewer_wnzZOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies bounded persistent memory for long-horizon text-to-SQL. It represents completed interactions as memory entries and learns P-UTL, an MLP that predicts whether retaining an entry will help or harm future SQL execution. Training labels come from counterfactual replay with and without individual retrieved memories; deployment retains the highest-scoring entries. The authors also construct extended SParC/CoSQL workloads with delayed repeated questions and user-defined conventions.

Critical strengths
  1. This paper studies an interesting problem about bounded text-to-SQL memory.

  2. Good results within a carefully controlled setting.

Critical weaknesses
  1. The evaluation does not sufficiently distinguish general memory utility from adaptation to a narrow generator.

  2. The supervision is not sufficiently validated as a surrogate for bounded-pool eviction.

  3. Lack necessary figure to describe the technical architecture.

Questions for authors
  1. Can a single fixed predictor be evaluated on held-out convention forms or recurrence patterns, against a baseline that identifies explicit definitions and deduplicates equivalent predicates using only the observed transcript? This would materially clarify whether the gains require learned execution utility rather than recognition of the generator's regular structure.

  2. Table 5 only reports the online manager cost. How many SQL calls are needed to train P-UTL-FULL, and what is the total training cost? How many deployment runs are needed to make up for this cost? Does changing the SQL answerer require relabeling?

3 comments
AuthorsOpenReview ↗

Response to Reviewer WnzZ (1)

We thank the reviewer for their valuable comments. We hope the following clarifications on generalization, definition-priority memory management, and computational cost analysis address the concerns.

1. Generalization to unseen convention forms and recurrence patterns

We agree that evaluating held-out convention forms and recurrence patterns would help distinguish generalization from adaptation to the trajectory generator. To address this concern, we evaluate a fixed P-UTL-FULL predictor under two distribution shifts and introduce a definition-priority baseline.

1. We switch the convention pattern by replacing the original lower-bound predicates with interval predicates and change the aggregation requested at the callback. For example, a original convention–callback pair looks like:

[CONVENTION] For the rest of our conversation, a 'large total_value_purchased' is a row in Product_Suppliers where total_value_purchased > 76000. Use this definition. To start: how many large total_value_purchased entries are there? 

[CALLBACK] Going back to our earlier definition: what is the SUM of total_value_purchased across all large total_value_purchased?

And a modified pair of it would be:

[CONVENTION] Use the term "large total_value_purchased" for records in Product_Suppliers satisfying "total_value_purchased" >= 76000 AND "total_value_purchased" <= 87000. How many large total_value_purchased entries are there? 

[CALLBACK] Going back to our earlier definition: what is the minimum value of total_value_purchased across all large total_value_purchased? 

The interval-based convention form is absent from training.

2. We change the recurrence patterns by sampling convention–callback gaps uniformly between 15 and 50 turns. Although these shorter gaps generally make memory retention easier, this recurrence distribution differs from that used during training.

3. We add an additional baseline called Def-priority, which extracts explicit definitions from the observed transcript, deduplicates their predicates, and assigns the corresponding entries the highest retention priority. When the memory pool reaches capacity, it applies a first-in, first-out eviction heuristic.

We use P-UTL-FULL, trained on the union of all training trajectories, without retraining for either distribution shift. Test trajectories use ($\mu=35$), callback$=5$, and the standard spacing of Scenario 3.

Evaluation conditionP-UTL-FULLE-LLMDef-priority
Original convention, original recurrence100.073.1100.0
OOD convention, original recurrence98.356.771.7
Original convention, OOD recurrence100.080.0100.0

Def-priority matches P-UTL-FULL under the original convention form, but its performance drops under the modified conventions. Under the recurrence shift, P-UTL-FULL and Def-priority maintain their performance, while E-LLM improves slightly. These results support P-UTL-FULL’s robustness to the tested distribution shifts, particularly the change in convention form.

In addition to this, we tested the different methods’ performance on harder trajectories, where the number of callbacks increases to 30 on the same original distribution setups.

Evaluation conditionP-UTL-FULLE-LLMDef-priority
Original convention, original recurrence94.266.769.2
OOD convention, original recurrence88.853.841.7
Original convention, OOD recurrence94.672.167.9

In this experiment we see a clear disadvantage of Def-priority on both OOD convention and OOD recurrence, while our method remains robust.

Appendix C.1 additionally compares against heuristic methods that retain episodic memories, which include definition-related information. These comparisons complement the definition-priority control by examining broader retention strategies. Together, the results highlight the importance of deciding not only which memories to admit, but also which to evict under a bounded memory budget.

AuthorsOpenReview ↗

Response to Reviewer WnzZ (2)

2. Training-data requirements and computational cost

Our training set contains 144 trajectories from Scenario 2 and 256 from Scenario 3. Constructing the training labels required 507,973 labeling calls. This is a one-time offline annotation cost for the reported training set, and therefore separated from the online memory-management cost reported in Table 5.

P-UTL-FULL is a three-layer predictor with 245,507 trainable parameters. Training the predictor on the annotated data takes less than one hour on a single NVIDIA A6000 GPU.

Regarding changes to the SQL answerer, the trajectory protocol does not require successful answers for subsequent rounds to proceed. As described in Appendix B.1, the correction protocol provides feedback for an initial bounded number of incorrect attempts, after which the interaction proceeds regardless of final answer correctness. Thus, the trajectory structure can be reused across SQL answerers.

To assess sensitivity to the amount of supervision, we additionally train P-UTL-FULL using different fractions of the training data and evaluate each predictor on Scenario 3:

Training fractionScenario 3 Performance
20%92.92
40%94.45
60%94.53
80%93.01
100%94.24

Performance remains within a relatively narrow range across the tested fractions. Using only 20% of the training data achieves 92.92 testing accuracy, compared with 94.24 using the full dataset. These results suggest that strong performance in this setting does not require the full annotation set.

Dear Reviewer,

While the rebuttal is ending soon, we looking forward to your response to check whether our rebuttal addressed your concerns. We would be happy to make clarifications if you had concerns unresolved.

Best, Submission 22100 Authors

Comments on the paper

1 comment

Dear Area Chairs and Reviewers,

We thank the reviewers for their feedback and for the additional experiments they suggested. As the rebuttal period draws to a close, we summarize the strengths identified by the reviewers, the concerns they raised, and the experiments and clarifications we provided in response.

Strengths

1. Framing long-horizon text-to-SQL as a bounded memory problem. Reviewers wnzZ and EGyU recognized the importance of deciding which past interactions to retain under a fixed memory budget. Reviewer EGyU emphasized that an early definition or business rule may become relevant again many turns later.

2. Lightweight predictor based on counterfactual utility. Reviewers CQHT and EGyU appreciated the use of counterfactual SQL execution to supervise retention decisions. Reviewer CQHT also highlighted the practical benefit of learning this signal offline. P-UTL updates the memory pool in approximately 2 ms without an online memory-manager API call.

3. Outstanding performance over controlled evaluations. Reviewer wnzZ recognized the results in a carefully controlled setting. Reviewer CQHT highlighted the comparisons on shared trajectories with varying callback delays, definition densities, and memory budgets.

Questions and Our Responses

1. Generalization beyond the trajectory generator. wnzZ asked about unseen conventions and recurrence schedules, CQHT requested a definition-priority baseline, and EGyU questioned reliance on templated wording. We tested a fixed predictor against Def-priority under both distribution shifts and separately evaluated free-form definitions (see sec. 1, Rebuttal to wnzZ and sec. 2 to EGyU). P-UTL** maintained accuracy** under the new schedule, outperformed Def-priority on unseen conventions, and surpassed all baselines on free-form definitions.

2. Harder eviction flows. EGyU asked what happens when new definitions exceed the memory budget. We introduced a new experiment that continuously introduced definitions and sampled callbacks from the full history (see sec. 1, Rebuttal to EGyU). P-UTL outperformed both E-LLM and LRU under this workload by a large margin.

3. Supervision for eviction. CQHT asked whether replacement-aware labels change downstream performance. We retrained P-UTL with replacement-aware labels (see sec. 2, Rebuttal to CQHT). Accuracy fluctuates, but does not reflect a significant performance change.

4. Memory capacity and retrieval. CQHT requested an unbounded-store comparison, an explanation for the accuracy drop with a larger pool, and a paired uncertainty estimate. We tested an unbounded store, inspected retrievals at matched rounds, and computed a paired bootstrap interval (see sec 3 & 4 to CQHT). Results showed that the unbounded variant worked well on simple setups but our method outperformed it when the memory pressure increased.

5. Offline cost and data requirements. CQHT requested offline costs, while wnzZ also asked about labeling requirements, amortization, and reuse across SQL answerers. We reported labeling and training costs, varied the training-set size, and clarified when trajectory structure can be reused (see sec. 5 CQHT and sec 2 to wnzZ).

Overall, we believe this paper merits acceptance because the key insight we wish to deliver is the value of predictive memory: a system with a fixed memory budget can learn which past interactions to retain by predicting their usefulness for future queries. Even with rich context and a strong backbone, a summarization-based method such as E-LLM still faces the decision of what to keep. Our lightweight predictor uses downstream SQL execution to guide that decision and outperforms all baselines in our long-horizon text-to-SQL experiments, particularly under memory pressure. We also provide a simulation framework that allows researchers to generate customized long trajectories with various constraints such as memory distance and memory pressure. This would benefit future works on long-horizon text-to-SQL.

We thank the reviewers for helping us sharpen the evaluation and will incorporate their feedback into the revised manuscript. In the meantime, we still welcome further feedback on whether we have addressed their concerns, and are happy to discuss.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.