- Rating
- 2
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
The paper separates persistent procedural Cards from task-specific Strategies for tool-using LLM agents. At execution stalls, a contextual-bandit controller either regenerates the Strategy while retaining the Card or retrieves a different Card. Completed tasks update Card usage statistics and can add new procedures. The evaluation reports a 3.00 percentage-point improvement over MemCon averaged across four interactive benchmarks, additional QA results, and ALFWorld analyses of controller choices, memory evolution, and token consumption. The proposed distinction is sensible, and the experiments provide encouraging evidence, but attribution of the gains to the specific lifecycle design remains incomplete.
Critical strengths
-
The separation between a reusable procedure and its current instantiation addresses a concrete failure mode: a locally unsuccessful plan need not invalidate the underlying procedural knowledge. Table 2(a) goes beyond a no-memory comparison by including fixed interventions, Coupled Refresh, non-contextual UCB, and LinTS. The 68.91% result for LinUCB versus 64.93% for Coupled Refresh is useful evidence that selecting the intervention level can matter.
-
The empirical coverage is reasonably broad. Table 1 reports improvements on all four interactive benchmarks with three evaluation runs, while the ALFWorld efficiency study includes three backbones. Table 2(b) separately examines memory growth and usage history. The paper also appropriately states that the longer Card lifetimes are descriptive consequences of the controller structure, rather than independent causal validation of the hierarchy (Section 5.3).
Critical weaknesses
-
The incremental benefit over existing adaptive memory control is not isolated cleanly enough. MemCon already learns context-dependent memory operations; the critical question is what the Card/Strategy distinction and segment-level feedback add under otherwise matched conditions. The internal ablations are helpful, but Coupled Refresh changes the intervention-selection rule as well as how the two levels are refreshed. Meanwhile, Appendix A specifies GPT-5.6 Sol for DUAL-MEM's Card extraction but does not establish an equivalent construction model and budget for each baseline. Sharing source trajectories alone does not eliminate this possible confound. A comparison using the same Cards, extractor, Watcher, execution prompts, and construction resources would make the modest full-suite gain much more persuasive.
-
Important parts of the control and continual-evaluation protocol remain ambiguous. Table 5 describes progress in terms such as a task-relevant state or a verified intermediate condition, without specifying the exact rules or prompts that implement those judgments. These signals determine both intervention timing and controller reward. In addition, Appendix B states that LinUCB parameters persist across tasks, whereas the slow-lifecycle experiment describes a fixed Fast Lifecycle and frozen probe state. It is unclear whether the controller matrices and reward vectors are identical across the four Table 2(b) variants, or whether some variants also benefit from additional controller learning on the 120-task stream. This distinction directly affects attribution of the 68.91% to 73.88% improvement to persistent-memory evolution.
Questions for authors
-
Clarification of the main comparison: Which model and construction budget produced each baseline's memory, especially MemCon's? Is there a matched comparison that changes only the lifecycle/controller design while sharing the extracted Cards and runtime components?
-
Clarification of Table 2(b): What happens to the LinUCB matrices and reward vectors during the 120-task stream and the held-out probe? Please specify the initialization, updates, and resets for every variant, and provide results with matched controller state if those states currently differ.
-
Clarification of the progress signal: Please provide the executable adapter rules or prompts underlying Table 5, including how task relevance and verification are determined in AppWorld, tau-bench, and GAIA. Evidence that these adapters use the same permitted observations and feedback across compared methods would materially strengthen the result.
We thank the reviewer for the thorough feedback and for acknowledging the value of our intervention ablations and the breadth of the evaluation. We appreciate the emphasis on experimental attribution and reproducibility, and we clarify the two comparison settings and the continual-evaluation protocol below.
R1. Baseline comparability and contribution of the two-level design (W1 / Q1)
Our paper uses two complementary comparisons. In the main six-benchmark study (Table 1 and Appendix A), memory-based methods share the benchmark-specific source trajectories, execution backbone, task order, tool interface, and evaluator. Each baseline retains its native memory representation and construction procedure. This evaluates complete systems under shared source experience and execution conditions. The interactive-benchmark average increases from 46.30% (MemCon) to 49.30% (DUAL-MEM); Appendix A, Table 4, also reports 44.05% (DUAL-MEM) versus 41.87% (ReMe) on AppWorld.
Table 2(a) then isolates the intervention-policy choice within a shared representation: all variants use the same frozen 60-Card Bank, retrieval index, Watcher, and Usage State. With these components fixed, contextual LinUCB reaches 68.91%, compared with 64.93% for Coupled Refresh. Our additional deterministic R→Q baseline achieves 64.68% under the same protocol. The new rule comparison reinforces the value of choosing the intervention level from execution context while holding the underlying procedural-memory system constant.
R2. Continual-evaluation protocol and Controller updates (W2 / Q2)
Section 4.2 and Appendix A specify a shared 60-Card initial Bank, a separate 120-task continual stream, and evaluation on the same 134-task held-out ALFWorld probe. Table 2(b) keeps the Fast Lifecycle design fixed and varies Bank Evolution and Usage History. The results show distinct contributions:
Appendix B (Eqs. 20–21) specifies LinUCB initialization (A_a = I₉, h_a = 0) and segment-level parameter updates. It also specifies that persistent memory updates are applied after task termination. We appreciate the request for an explicit variant-by-variant account of the Controller state and will clarify when the LinUCB matrices and reward vectors are initialized, updated, or frozen during the continual stream and held-out probe. This will make the Table 2(b) protocol easier to reproduce and interpret.
R3. Observable progress and benchmark adapters (W2 / Q3)
Appendix B, Eq. (17), defines segment progress as a binary event: Δp = 1 when at least one observable task-relevant progress event occurs after intervention. Table 5 identifies the evidence used for each environment:
The binary definition gives the Controller a common segment-level feedback scale across environments with different native signals. As described in Appendix B, adapters use observations available through the environment or tool interface. Segment utility then combines progress with completion, intervention cost, and stagnation (Eqs. 18–19). We will expand the adapter descriptions with precise conditions and examples, especially for AppWorld, τ-bench, and GAIA, as suggested.
Thank you again for these detailed comments. They have helped us improve the clarity of the experimental protocol and the presentation of the controlled evidence.