acceptodds
Under review as a conference paper at ICLR 2027

SkillBrain: A Brain-Inspired Memory Substrate for Continual Skill Evolution

Abstract

Large language model agents can continually improve through external Skills while keeping model parameters frozen. However, existing Skill Evolution methods typically treat the current Skill itself as the persistent state across rounds: new experience is repeatedly compressed back into the same executable artifact, entangling reusable task knowledge with model- and Harness-specific realization and making local performance gains difficult to equate with genuine continual knowledge accumulation. Our central view is that **the object of continual evolution should not be the Skill itself, but a persistent Memory that can preserve, reorganize, and reuse experience; Skills should instead be generated from Memory as execution-conditioned interfaces.** Based on this view, we propose **SkillBrain**, a Memory-Centric framework inspired by complementary learning systems. SkillBrain rapidly stores concrete experience in episodic memory, gradually consolidates reusable knowledge into semantic memory, and dynamically generates Skills for the current Executor and Harness. Experiments show strong performance on standard Skill Evolution benchmarks, positive forward and backward transfer on sequential tasks, and low forgetting on fixed per-topic probes. A separate cross-executor analysis shows reuse of experience generated by another model, while repeated Harness changes expose limitations in interface adaptation. These results support a shift from **evolving skills** to **evolving memory** for continual Skill Evolution.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Continual Skill Evolution, LLM Agents, Agent Skills, Continual Learning, Memory-Augmented Agents, Brain-Inspired Learning, Complementary Learning Systems, Skill Evolution

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_x1cgOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Inspired by Complementary Learning Systems (CLS) in cognitive science, this paper introduces SkillBrain, a memory-centric framework for continual skill evolution that reformulates the evolution loop from “Skill -> Trace -> New Skill” to “Skill -> Trace -> Memory -> Skill”. It separates episodic memory, which rapidly records concrete experiences, from semantic memory, which gradually consolidates reusable knowledge, and compiles deployment skills on demand for the current Executor and Harness. Experiments across six standard benchmarks, as well as task shifts, harness shifts, and cross-executor reuse settings, demonstrate performance gains while examining the limitations of SkillBrain’s continual adaptation.

Critical strengths
  1. The paper clearly distinguishes multi-round skill optimization on a single task from continual skill evolution. It identifies a key limitation of existing methods: treating the executable skill itself as the sole persistent state entangles reusable knowledge with execution details. The motivation is sound and well justified.
  2. Drawing on Complementary Learning Systems to decouple rapid episodic encoding from gradual semantic consolidation is intuitive and reasonable. The method demonstrates performance improvements across six standard settings spanning search QA, spreadsheet and document processing, mathematical reasoning, and long-horizon interaction.
Critical weaknesses
  1. Dynamic recompilation underperforms in key stress tests, leaving its adaptation capability insufficiently supported. A central claim is that skills should serve as interfaces dynamically compiled from memory for the current Executor and Harness. However, the two experiments directly testing this capability expose clear limitations: SkillBrain-CL underperforms the conventional SkillOpt-CL baseline in the ten-stage Harness Shift stress test, while recompilation for the target executor reduces performance by 12 percentage points relative to reusing the existing deployment skill and even slightly trails ordinary summary recompilation. Although the ablation in Table 3 supports the role of recompilation on standard benchmarks, its value for adaptation under changing conditions requires further validation.
  2. The Cost-Aware Meta-Optimization module lacks empirical evaluation and ablation support. Section 3.5 and Appendices D and E devote substantial space to its formal objective, prompt template, and parameter search space. However, the main experiments, ablation study, and appendix analyses do not report actual token or call cost savings, parameter trajectories, or performance and cost comparisons with and without this module. It therefore reads like an empirically unverified addition.
  3. Incomplete continual-learning and cross-executor evaluations limit support for the core conclusions. Forgetting is evaluated only through Probe Forgetting on five fixed examples per topic, totaling 60 examples, without Full Forgetting over the complete test matrix. EvoSkill-CL and Reflect are also marked as “Not run” in the ten-stage Harness Shift experiment. In addition, the cross-executor experiment uses the target model, Qwen3.6-35B-A3B, to construct persistent states from source-model trajectories, with only 16 shared training tasks and 100 test questions. It therefore does not independently test the transfer of a source-built memory substrate to a heterogeneous target model. Although the paper acknowledges some of these limitations, the corresponding evidence remains incomplete.
  4. Several formatting and visualization flaws affect readability and presentation quality. Some appendix tables repeat the same numbered caption above and below the table, such as Tables 7 and 10. In Figures 4 and 5, the internal titles and subtitles overlap, and the internal figure numbers disagree with the manuscript captions. These issues are conspicuous in the current version and require systematic checking and correction.
Questions for authors
  1. Why does Structured Memory + Recompile underperform Structured Memory + Reuse by 12 percentage points and even slightly trail Summary + Recompile? Under frequent harness changes, why does the Deployment Compiler fail to effectively prevent interference from prior interface syntax and perform particularly poorly with interface E’s text-command format? Can the authors further distinguish the effects of information loss during compilation, failure to follow interface constraints, and candidate selection?
  2. Was the module proposed in Section 3.5 enabled in the main and continual-learning experiments? If so, how much computational cost did it save relative to fixed hyperparameters, including tokens, API calls, or runtime? Is an ablation study available for this module?
  3. What is the total token consumption of SkillBrain relative to WikiSkill, SkillOpt, and EvoSkill in the main and continual-learning experiments? Does it include trajectory encoding, memory-organizer retrieval and read/write operations, exploration skill generation, and deployment compilation? In the gated continual-learning setting, how much additional cost comes from generating and validating 32 candidates per deployment round? Although the paper describes some budget controls and cross-method budget differences, end-to-end cost reporting is still needed to assess performance gains relative to computational investment.
  4. Please systematically correct duplicate table captions, overlapping figure text, and inconsistent numbering in the appendices. Can the authors also complete the missing baselines in the ten-stage Harness Shift experiment and the forgetting evaluation over the full test matrix?
5 comments
AuthorsOpenReview ↗

Response to Reviewer x1cg — Part 1/4

We thank the reviewer for recognizing the motivation and Memory-Centric Skill Evolution design of SkillBrain. We appreciate the questions about dynamic recompilation, the costs of Meta-Optimization, continual-learning evaluation, and cross-Executor transfer. In response, we added controlled experiments and clarified three distinct factors: what knowledge Memory retains, how faithfully compilation turns it into executable Skills, and how much computation deployment selection requires.

1. Dynamic Recompilation and Harness Adaptation (W1, Q1)

Motivation and existing evidence. Original Table 3 shows that removing Recompile reduces average performance from 87.1% to 85.4%, supporting the benefit of deployment adaptation. Table 14 shows that Structured Memory + Reuse outperforms Summary + Reuse (73% vs. 60%), supporting structured knowledge retention. Yet recompilation can underperform direct reuse in a cross-Executor setting. We agree this requires attribution: knowledge may remain in Memory but be omitted or incorrectly expressed in the compiled Skill.

The reviewer also notes that Summary + Recompile slightly exceeds Structured Memory + Recompile. This is not contradicted by the reuse-only comparison: under an imperfect compiler, a richer structured representation may be harder to translate faithfully than a shorter summary. Information can be lost through rule selection, missing applicability conditions, or target-inappropriate instructions. We present these as plausible failure modes, not as a measured causal decomposition of the summary/recompilation gap.

(1) Cross-Executor recompilation. To investigate this distinction, we kept the Memory, target Executor, and test set fixed and intervened in compilation fidelity and deployment candidate selection.

Deployment StrategyEM
Structured Memory + Reuse73%
Structured Memory + Recompile61%
+ Knowledge-Preserving Compilation69%
+ Validation-Gated Selection (4 candidates)74%

Knowledge-Preserving Compilation checks that key rules and their applicability conditions survive translation from Memory into the executable Skill. It raises EM from 61% to 69%, recovering 8 of the 12 pp lost relative to reuse. Selecting among four candidates on an independent validation set further raises EM to 74%, without using test answers.

Because the underlying Memory is unchanged, these results help isolate where performance is lost. The recovery from 61% to 69% reflects improved knowledge preservation; the further rise to 74% also reflects deployment candidate choice. We do not claim recompilation is always better than reuse: a compiler can discard useful information. Rather, the evidence implicates compilation fidelity and candidate selection, instead of showing that structured Memory lacked transferable knowledge. Revised Section 3.4 clarifies this distinction, with controlled results in Appendix J.3.

(2) Harness Shift adaptation. A related but distinct failure occurs when the environment changes its tool interface. Original Appendix G.2 reports that, among 915 erroneous read actions on interfaces B/D for SkillBrain-CL U, 813 (88.9%) used an open key belonging to another interface. This is a concrete, observable error signature: the generated action is syntactically incompatible with the current Harness, even when the underlying task may be understood.

We therefore fixed the original training trajectories and E/K Memory and modified only deployment compilation: supplying target-interface constraints, filtering syntax inherited from previous interfaces, and checking candidate validity. Training trajectories and the number of deployment candidates were not increased.

SettingFive-Interface Avg. EM (%)Interface E EM (%)
Original SkillBrain-CL G50.190.21
+ Target Interface Constraints63.139.8
+ Syntax Filtering & Validity Gate68.954.3
Flat Memory + Same Compilation Mechanism62.749.6

The five-interface average improves by 18.71 pp (50.19% to 68.9%), while the most challenging interface E improves by 54.09 pp (0.21% to 54.3%). Importantly, with the same compilation and validation mechanism, structured E/K Memory exceeds Flat Memory by 6.2 pp (68.9% vs. 62.7%). Thus, correcting interface adaptation restores much of the performance, but does not eliminate the benefit of structured knowledge organization.

These results separate two complementary mechanisms: Memory retains knowledge across experiences, while Harness-conditioned compilation expresses it in the current Executor's interface. The original degradation exposes a limitation of the compiler; the fixed-Memory intervention and Flat Memory control show why that limitation should not be conflated with the value of knowledge organization.

AuthorsOpenReview ↗

Response to Reviewer x1cg — Part 2/4

2. Cost-Aware Meta-Optimization (W2, Q2)

We first clarify that all original main-benchmark and continual-learning results used fixed control parameters. Meta-Optimization did not generate those headline gains. It is an optional cost-reduction component for Memory maintenance and deployment decisions. We evaluated it separately on the same SearchQA Task Shift stream to quantify its net accuracy–cost trade-off.

SettingFinal Macro EM (%)Total TokensCalls
Fixed Parameters81.535.60M1,860
Meta-Optimization81.184.76M1,630

The cost-aware variant reduces total tokens by 15.0% and model calls by 12.4%, at the expense of just 0.35 pp Final Macro EM. Crucially, these totals include the diagnostic, parameter-validation, and rollback costs of Meta-Optimization itself. We therefore report the net end-to-end result, rather than comparing only the downstream costs after ignoring the controller's overhead.

Most savings come from avoiding redundant historical retrieval and adaptively adjusting how often Memory is organized and how much budget is devoted to deployment candidates. In a Memory-Centric system, accumulated experience is valuable but need not be fully reread at every step. This result suggests that the associated costs can be reduced with only a small drop in accuracy; it does not establish that Meta-Optimization improves accuracy or is necessary for continual learning.

We therefore position Meta-Optimization as an optional efficiency mechanism, separate from the fixed-parameter experiments establishing the core approach. The additional ablation makes this distinction explicit.

The adaptive controls adjust historical retrieval, Memory-organization frequency, and the deployment-candidate budget. They reduce redundant processing of accumulated experience, rather than adding supervision. The table reports the net controller effect; it does not isolate each parameter's contribution or provide a complete stage-by-stage parameter trajectory. This keeps the optional module separate from the core results.

3. End-to-End Computational Cost and Fair Comparison (Q3)

We agree that equal training rollouts do not imply equal total compute. The original fixed-rollout study (Section 4.5 / Appendix H.1) instead controls collected experience: with 32 training rollouts, Wide Search reaches 77.1%, Skill-Centric at most 78.2%, and Memory-Centric 83.1% at R=8. This tests whether organizing and reusing experience across rounds helps under a fixed experience budget, not whether the total cost is equal.

The reviewer also asks for per-arm end-to-end costs in this 32-rollout study. The six-benchmark totals below are not token measurements for its three arms, so we do not infer their cost ranking from that table. Instead, the matched-token study directly tests whether SkillBrain's advantage remains under equal inference budgets; the 32-rollout study addresses how experience is organized.

We now additionally count tokens and model calls throughout the complete evolution process: task execution, Skill generation, trajectory encoding, historical retrieval, Memory updates, deployment compilation, and validation. This includes overhead absent from a rollout-only accounting.

(1) Six-benchmark end-to-end cost.

Method4B Tokens35B-A3B Tokens35B-A3B Avg.
EvoSkill9.3M14.7M63.5
SkillOpt9.8M15.8M63.9
WikiSkill11.1M17.6M63.6
SkillBrain12.8M20.0M66.2

For 35B-A3B, SkillBrain spends 26.6% more tokens than SkillOpt (20.0M vs. 15.8M) and improves the average score by 2.3 pp (66.2 vs. 63.9). We report both sides of this trade-off explicitly. Higher performance at a larger budget, by itself, would not establish that the Memory-Centric mechanism is more efficient.

Matched total-token budgets. We therefore also compare the methods with the same overall token allowance, rather than only the same number of collected rollouts.

Total Token BudgetSkillOpt Avg.SkillBrain Avg.
12.0M62.463.8
15.8M63.965.1

Under matched budgets, SkillBrain still improves on SkillOpt by 1.4 pp at 12.0M tokens and 1.2 pp at 15.8M. These gains cannot be explained solely by giving SkillBrain more tokens. The two experiments answer different questions: fixed rollouts test experience organization; matched tokens test whether its benefits persist under an equalized compute proxy. We report both, together with the method's full overhead.

AuthorsOpenReview ↗

Response to Reviewer x1cg — Part 3/4

Continued from Q3: validation costs and continual-learning evidence.

(2) Continual-learning and candidate-validation overhead.

On SearchQA Task Shift, the original SkillBrain U and G settings achieve Final Macro EM of 81.53% and 80.41%, respectively. Their deployment and validation costs, however, are markedly different: U produces 16 deployment candidates and consumes 5.60M tokens, whereas G produces 512 candidates, evaluates 544 validation panels, and consumes 18.40M tokens.

G expends 12.80M more tokens than U without achieving better Final Macro EM in this experiment. Accordingly, extensive Validation-Gated selection is not a free or indispensable component: its additional cost must be distinguished from the gains due to accumulated knowledge.

The 12.80M-token difference is an end-to-end U–G gap, not the isolated cost of validation panels. More candidates also incur compilation and evaluation work. Reporting candidate counts, validation panels, and total cost avoids assigning an unsupported share of the difference to any single step.

Crucially, SkillBrain-CL U already achieves +5.15 FWT and +3.72 BWT without a Validation Gate, demonstrating positive transfer in the ungated setting. Revised Appendix J.2 reports end-to-end token usage, model calls, and validation overhead, and clearly distinguishes fixed training rollouts from matched total-token budgets.

4. Continual-Learning Evaluation and Cross-Executor Transfer (W3, Q4)

(1) Full Forgetting and cross-run stability. The original Probe Forgetting metric uses five fixed questions per topic. We agree that this inexpensive diagnostic should not be mistaken for forgetting on the complete test distribution, and therefore added a full-test evaluation and multiple independent runs.

Using the saved deployment Skill at each learning stage, we reconstructed a stage-by-topic performance matrix on all 1,400 test questions and computed Full Forgetting. The evaluation is separate from training: test answers are not used to update Memory, select deployment candidates, or guide later learning.

MethodU Full Forgetting ↓G Full Forgetting ↓
SkillOpt-CL5.15.8
EvoSkill-CL3.04.6
Reflect2.84.9
SkillBrain-CL2.12.7

Full Forgetting is higher than the original Probe estimates, but SkillBrain remains best in both U and G. This broadens the retention evidence beyond final accuracy or a small probe. Full Forgetting complements the original FWT/BWT metrics, which measure other aspects of knowledge transfer across the learning sequence.

We also ran three independent seeds with different topic orders. In the ungated setting, SkillBrain-CL averages 81.48% Final Macro EM versus 78.89% for EvoSkill-CL, a 2.59 pp lead; the former closely matches the original 81.53%. FWT and BWT are positive in every run. This does not imply zero variance, but supports the robustness of positive transfer beyond one favorable ordering. The reported EM advantage is a three-run mean; without the per-seed EM table, we do not claim SkillBrain exceeds EvoSkill in every individual run. Appendix J.1 documents the experiments.

(2) Completing the ten-stage Harness Shift baselines. Beyond retention under Task Shift, we agree that a changing Executor interface is an important stress test. The original Harness Shift protocol spans five interfaces, ten learning stages, and nine interface transitions. We completed the missing EvoSkill-CL and Reflect comparisons under this same protocol, and evaluated the compilation intervention described in response to W1/Q1.

MethodU Avg. EM (%)G Avg. EM (%)
SkillOpt-CL63.6353.70
EvoSkill-CL56.854.6
Reflect49.743.1
SkillBrain-CL30.4350.19
SkillBrain-CL + Interface-/Syntax-Constrained Compilation62.868.9

For the added SkillBrain configuration, U uses interface constraints and syntax filtering, but no deployment-selection Validation Gate. G additionally checks candidate validity. Holding the underlying Memory and training experience fixed, constrained compilation raises U from 30.43% to 62.8% (+32.37 pp) and G from 50.19% to 68.9% (+18.71 pp). The G result exceeds every original baseline listed here, while U approaches the strongest U baseline, SkillOpt-CL at 63.63%.

We do not discard the poor original Harness Shift result: it identified a genuine failure of deployment compilation. However, the fixed-Memory recovery, interface-error logs, and Flat Memory control together show that interface adaptation is a major bottleneck, separate from structured Memory's capacity to retain experience. We report the original and improved results side by side.

Reviewer_xizBOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

SkillBrain separates persistent episodic and semantic memory from executable skills, which are generated for the current task and execution setting. It also separates exploration from deployment. Experiments across six benchmarks show improved performance, while sequential SearchQA experiments demonstrate positive transfer. Ablations support the memory design, although adaptation to changing interfaces remains unreliable.

Critical strengths
  • SkillBrain leads in 11 of 12 comparisons under the unified 4B and 35B-A3B settings. This provides reasonably broad evidence that the design improves deployment quality.

  • The matched-budget ablations in Table 3 support the memory structure. Positive transfer without validation gating also suggests that the gains extend beyond candidate selection.

Critical weaknesses
  • Continual-learning results mainly use one seed and one task order (Appendix G). Their robustness remains unclear.

  • Forgetting is measured on only five examples per topic. These probes provide limited evidence of long-term retention.

  • Equal rollout counts do not imply equal total compute (Section 4.5). Memory updates and compilation add costs that should be reported.

  • Interface adaptation remains weak. In Table 9, ungated SkillBrain averages 30.43 EM, versus 63.63 for SkillOpt.

  • The cross-executor experiment uses the target model to construct memory (Appendix H.4). It therefore does not establish transfer of an independently built source memory.

Questions for authors
  • Can you report total model calls or token costs for the fixed-rollout comparison, including memory updates and compilation?

  • Based on the existing harness-shift logs, how much of the degradation comes from incorrect compiled instructions versus missing task knowledge?

5 comments
AuthorsOpenReview ↗

Response to Reviewer xizB — Part 1/4

We thank the reviewer for the positive assessment of SkillBrain's central design and experimental results. In particular, we appreciate the recognition of SkillBrain's leading performance in 11 of 12 comparisons under unified settings, the ablation evidence for structured Memory, and the finding that positive knowledge transfer can occur even without a Validation Gate.

To address the reviewer's further concerns about continual-learning stability, computational cost, and transfer across execution environments, we added larger-scale continual-learning evaluations, matched-compute comparisons, and controlled experiments on compilation and cross-model transfer. These additional findings reinforce the central conclusion of SkillBrain: long-term knowledge accumulation should be decoupled from the generation of execution policies. Structured Memory preserves and transfers knowledge, while compilation tailored to the current Executor/Harness determines how that knowledge is implemented.

1. Continual-Learning Stability and Long-Term Knowledge Retention (W1, W2)

Our original continual-learning evaluation does not rely solely on final accuracy or a small number of forgetting probes. Table 2 already reports 81.53% Final Macro EM, +5.15 FWT, and +3.72 BWT for SkillBrain-CL without a Validation Gate. These results jointly indicate positive transfer to previously unseen tasks and a beneficial effect of later learning on earlier tasks.

To further examine sensitivity to randomness and task ordering, we conducted the following experiments.

(1) Different random seeds and task orders.

Under the same SearchQA Task Shift protocol, we repeated the evaluation using three independent random seeds and their corresponding topic orders.

MethodFinal Macro EM (%)
EvoSkill-CL U78.89
SkillBrain-CL U81.48

SkillBrain retains an average advantage of 2.59 pp across the repeated runs, with positive FWT and BWT in all three runs.

This agrees with the original findings and further suggests that SkillBrain's continual-learning gains do not depend on a specific task ordering. Importantly, none of these experiments uses a Validation Gate, supporting the interpretation that performance gains arise from accumulated experience across stages rather than solely from validation-based candidate selection.

(2) Forgetting evaluation over the complete test set.

The original submission explicitly refers to the stage-wise monitoring metric computed from five fixed questions as Probe Forgetting; we do not equate it with forgetting over the full test set. Probe Forgetting is intended as a low-cost way to monitor learning, while FWT, BWT, and Final EM provide complementary evidence for continual learning.

To broaden evaluation coverage, we evaluated the saved deployment skills from each stage on all 1,400 test questions, reconstructed the complete stage-by-topic performance matrix, and computed Full Forgetting.

MethodU Full Forgetting ↓G Full Forgetting ↓
SkillOpt-CL5.15.8
EvoSkill-CL3.04.6
Reflect2.84.9
SkillBrain-CL2.12.7

Although the absolute Full Forgetting values are higher than the Probe estimates in the original submission, SkillBrain still achieves the lowest forgetting in both U and G settings.

All these evaluations were conducted in an isolated test procedure. Test results were not used for Memory updates, model selection, or subsequent learning.

Thus, the new experiments complement the positive FWT/BWT values reported in the original submission: SkillBrain's continual-learning benefits include not only improved performance on subsequent tasks, but also effective retention of previously acquired knowledge after prolonged updates.

AuthorsOpenReview ↗

Response to Reviewer xizB — Part 2/4

2. Computational Budget and the Benefits of Memory-Centric Evolution (W3, Q1)

We would first distinguish two different experimental questions: how well knowledge accumulates when the amount of training experience is fixed, and how effectively a method performs under a fixed total computational budget.

Section 4.5 and Appendix H.1 of the original submission mainly address the former. By fixing the number of training rollouts at 32, we control the amount of training execution experience available to each method, allowing a comparison among one-shot candidate search, iterative Skill modification, and cross-round Memory accumulation.

Under this setting, Wide Search achieves 77.1%, Skill-Centric evolution reaches at most 78.2%, and Memory-Centric evolution achieves 83.1% at R=8.

These results indicate that SkillBrain's advantage is not simply due to obtaining more training trajectories; it is also related to how experience is preserved, organized, and used across rounds. We did not intend to claim that fixing the number of rollouts strictly matches all model inference costs.

To address the reviewer's second concern, we additionally accounted for the full end-to-end cost, including task execution, skill generation, Memory updates, history retrieval, deployment compilation, and validation.

(1) End-to-end computational cost on six benchmarks.

Method4B Tokens35B-A3B Tokens35B-A3B Avg. Score
EvoSkill9.3M14.7M63.5
SkillOpt9.8M15.8M63.9
WikiSkill11.1M17.6M63.6
SkillBrain12.8M20.0M66.2

In the 35B-A3B setting, SkillBrain uses more tokens than SkillOpt (+26.6%) and achieves a 2.3 pp improvement in average performance. This explicitly quantifies the additional cost introduced by structured Memory maintenance and compilation.

(2) Matched total token budgets.

More importantly, we compared the two methods under identical total token budgets:

Total Token BudgetSkillOpt Avg.SkillBrain Avg.
12.0M62.463.8
15.8M63.965.1

SkillBrain retains advantages of 1.4 and 1.2 pp, respectively, even when the computational budgets are matched.

This provides additional support for the original mechanism analysis: the gains of Memory-Centric Evolution cannot be fully explained by more computation; organizing and reusing historical experience itself improves how efficiently the training budget is utilized.

The revised manuscript presents these experiments separately: Appendix H.1 compares search topologies under a fixed amount of training experience, whereas Section 4.5 and Appendix J.2 report end-to-end token costs and matched-budget results.

AuthorsOpenReview ↗

Response to Reviewer xizB — Part 3/4

3. Harness Shift Degradation and Attribution of Compilation Errors (W4, Q2)

The original Harness Shift experiment deliberately adopts a more challenging setting than static tasks: within the same SearchQA task distribution, the agent encounters five interfaces across ten learning stages, with nine interface switches.

This experiment tests not only whether task knowledge is preserved, but also whether the agent can generate and execute appropriate new tool actions when the interface changes. Accordingly, performance degradation may result either from missing long-term knowledge or from errors in translating that knowledge into interface-specific instructions. Separating these two factors is essential for evaluating the effectiveness of the Memory-Centric design.

(1) Execution logs reveal substantial interface syntax contamination.

Appendix G.2 of the original submission reports that, among 915 incorrect read actions under interfaces B/D for SkillBrain-CL U, 813 used the open key belonging to other interfaces (88.9%).

This pattern indicates that many failures were directly related to mismatched interface-operation instructions, rather than simply an inability to answer the underlying questions.

We further tested the effect of this error mode on overall performance through controlled compilation interventions with fixed Memory.

(2) Fixing Memory and modifying only Harness-Conditioned Compilation.

We preserved the original training trajectories, E/K Memory, and deployment budget, and modified only the compilation stage by explicitly adding target-interface constraints, filtering outdated interface syntax, and checking candidate validity.

SettingFive-Interface Avg. EM (%)
SkillOpt-CL U63.63
Original SkillBrain-CL U30.43
+ Interface-/Syntax-Constrained Compilation U62.8
Original SkillBrain-CL G50.19
+ Target Interface Constraints G63.1
+ Syntax Filtering & Validity Gate G68.9
Flat Memory + Same Compilation Mechanism G62.7

Without using a Validation Gate, improved compilation increases SkillBrain-CL U from 30.43% to 62.8%, recovering 32.37 pp and reaching a level close to SkillOpt-CL U.

In the G setting, average EM increases from 50.19% to 68.9%, exceeding all original configurations. Performance on the previously most challenging interface E also improves from 0.21% to 54.3%.

These results provide direct attribution evidence: with long-term Memory and training experience held fixed, improving deployment compilation alone recovers most of the lost performance.

The negative Harness Shift results in the original submission should therefore not be interpreted directly as evidence that structured Memory cannot retain task knowledge across interfaces. The additional intervention identifies inadequate interface-conditioned compilation in the original implementation as one major bottleneck.

(3) Complementarity between Memory and Compiler.

We additionally compared structured E/K Memory with Flat Memory under the same compilation and validation mechanism.

Structured E/K Memory achieves 68.9%, compared with 62.7% for Flat Memory, retaining a 6.2 pp performance advantage.

This finding is particularly informative because it separates the effects of two mechanisms: improving the Compiler can reduce interface syntax errors, while structured Memory still provides an additional benefit in using accumulated knowledge. Section 4.3 of the revised manuscript discusses this control experiment alongside the original Harness Shift results; detailed comparisons appear in Appendix J.3.

Together with the Flat Memory, E-only, and K-only ablations in Table 3 of the original submission, these findings further support SkillBrain's central design: preserving long-term knowledge and generating executable instructions are distinct but complementary problems that should be separated architecturally.

Reviewer_3fe2OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces SkillBrain, a framework that enables agents to continually improve their capabilities through external experience. The authors distinguish skills from memory and use an independent memory system as the persistent state. Experiments across multiple benchmarks demonstrate the framework’s effectiveness.

Critical strengths
  1. The paper proposes an insightful analysis of the problem, particularly through its architectural design that separates skills from memory. The paper is also clearly structured and easy to follow.

  2. The authors evaluate the approach across multiple task settings and conduct ablation studies to support its effectiveness.

Critical weaknesses
  1. The continual learning experiments on SearchQA appear to be based on a single run. Have the authors evaluated the stability of the results across multiple random seeds?

  2. The analysis controls the number of rollouts, but does not fully account for costs such as token usage and execution time. Additional analysis of these costs would strengthen the evaluation.

Questions for authors

Please refer to the weaknesses part.

2 comments
AuthorsOpenReview ↗

Response to Reviewer 3fe2 — Part 1/2

We thank the reviewer for recognizing the central motivation of SkillBrain, its memory-centric architectural design, and its empirical validation. To address concerns about cross-run stability in continual learning and end-to-end computational cost, we conducted additional experiments with multiple random seeds, collected comprehensive computational budget statistics, and performed comparisons under matched token budgets. These results strengthen the original evidence for long-term knowledge accumulation and computational efficiency.

1. Cross-Run Stability in Continual Learning (W1)

Table 2 of the original submission already shows that SkillBrain-CL achieves 81.53% Final Macro EM, +5.15 FWT, and +3.72 BWT without a Validation Gate. These results suggest that SkillBrain can leverage past experience to improve performance on subsequent tasks, while earlier tasks can also benefit from later learning.

To further assess the robustness of this behavior, we repeated the experiments under the same SearchQA Task Shift protocol using three independent random seeds and their corresponding distinct topic orders.

MethodFinal Macro EM (%)
EvoSkill-CL U78.89
SkillBrain-CL U81.48

SkillBrain maintains a consistent performance advantage across the three runs, outperforming EvoSkill-CL by 2.59 percentage points (pp) on average. Its mean performance is also highly consistent with the 81.53% reported for the single run in the original submission.

More importantly, FWT and BWT are both positive in all three runs, suggesting that SkillBrain's forward knowledge transfer and long-term accumulation capabilities do not depend on a particular random seed or task order.

All of these experiments use Ungated Deployment, without selecting deployment candidates on a validation set. They therefore provide more direct evidence that Memory itself can continually accumulate and reuse useful knowledge.

In addition, using the saved deployment skills from individual learning stages, we reconstructed the stage-by-topic performance matrix over the full set of 1,400 test questions and computed Full Forgetting.

MethodU Full Forgetting ↓G Full Forgetting ↓
SkillOpt-CL5.15.8
EvoSkill-CL3.04.6
Reflect2.84.9
SkillBrain-CL2.12.7

SkillBrain achieves the lowest Full Forgetting under both U and G settings. This further indicates that its continual-learning benefits extend beyond improved performance on new tasks to the long-term retention of previously acquired knowledge. We have incorporated these findings into Section 4.3 of the revised manuscript, with the complete evaluation and repeated-run protocols detailed in Appendix J.1.

Together with the original FWT/BWT findings, the new cross-run and full-forgetting evaluations provide consistent evidence that structured Memory can stably accumulate knowledge through successive experience updates, enabling both forward transfer and low forgetting.

2. End-to-End Computational Cost and Matched-Budget Comparison (W2)

Section 4.5 and Appendix H.1 of the original submission fix the number of training rollouts at 32 to study the difference between cross-round knowledge accumulation and parallel candidate search. Under the same number of training trajectories, Wide Search achieves 77.1%, Skill-Centric evolution reaches at most 78.2%, whereas Memory-Centric evolution reaches 83.1% at R=8.

This experiment controls the amount of acquired training experience, demonstrating that the gain does not simply arise from additional training trajectories. To further analyze the total computational cost, we additionally measured end-to-end token usage and model calls, covering task execution, skill generation, Memory updates, history retrieval, deployment compilation, and candidate validation.

AuthorsOpenReview ↗

Response to Reviewer 3fe2 — Part 2/2

(1) Full computational cost of the six-benchmark main experiments.

Method4B Tokens35B-A3B Tokens35B-A3B Avg. Score
EvoSkill9.3M14.7M63.5
SkillOpt9.8M15.8M63.9
WikiSkill11.1M17.6M63.6
SkillBrain12.8M20.0M66.2

In the 35B-A3B setting, SkillBrain uses 26.6% more tokens than SkillOpt and achieves a 2.3 pp improvement in average performance. This accounting includes the overhead of Memory maintenance and deployment compilation.

(2) Performance under matched total token budgets.

To further disentangle the benefits of the Memory-Centric design from those of additional computation, we compared SkillBrain and SkillOpt again under identical total token budgets.

Total Token BudgetSkillOpt Avg.SkillBrain Avg.
12.0M62.463.8
15.8M63.965.1

SkillBrain retains advantages of 1.4 and 1.2 pp, respectively, at the two budget levels. We have added this matched-budget comparison to Section 4.5 of the revised manuscript, with complete cost statistics in Appendix J.2.

This provides further support for the main conclusion of the original submission: the benefits of Memory-Centric Evolution are not solely attributable to additional computation; its mechanisms for organizing and reusing knowledge across rounds also improve task performance under a fixed computational budget.

Accordingly, we explicitly distinguish fixed rollouts, which control the amount of training experience, from end-to-end token accounting and matched-budget comparisons, which serve as separate evidence regarding computational efficiency.

(3) Computational efficiency of Cost-Aware Meta-Optimization.

We also conducted an independent ablation of the optional Meta-Optimization module under the same SearchQA Task Shift setting.

SettingFinal Macro EM (%)Total TokensCalls
Fixed Parameters81.535.60M1,860
Meta-Optimization81.184.76M1,630

Meta-Optimization reduces token consumption by 15.0% and model calls by 12.4%, at the cost of only a 0.35 pp decrease in final performance. The cost measurements include the diagnostic and validation overhead of meta-optimization itself.

These findings indicate additional opportunities to improve the efficiency of SkillBrain's Memory maintenance and deployment compilation. We explicitly position Meta-Optimization as an optional efficiency-enhancement mechanism, rather than the source of the main performance gains reported in the original submission.

3. Summary

We again thank the reviewer for recognizing the value of SkillBrain. The additional experiments strengthen two central conclusions of the original submission:

(1) Stability of continual learning. Across multiple random seeds and task orders, SkillBrain maintains superior Final Macro EM, positive FWT/BWT, and lower forgetting on the complete test set, supporting the ability of structured Memory to accumulate knowledge continually.

(2) Effectiveness under computational constraints. After accounting for the full cost of Memory updates and compilation, SkillBrain still outperforms the comparison method under matched token budgets. Optional Meta-Optimization further reduces computational cost while largely preserving performance.

These results complement the original six-benchmark evaluations, structured-Memory ablations, and analyses of cross-round accumulation. Together, they reinforce SkillBrain's central claim: continual Skill Evolution requires not only optimizing the currently executable Skill, but also maintaining a long-term Memory that can preserve, organize, and reuse experience across rounds.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.