acceptodds
Under review as a conference paper at ICLR 2027

STRATA: Stratified Trajectory-to-Knowledge Augmentation for Terminal Agents

Abstract

Terminal agents are becoming an important interface for software development, system administration, and complex environment automation. However, they still struggle with long-horizon terminal tasks where small command mistakes can change the environment and lead to repeated failures. Supervised fine-tuning and reinforcement learning can improve such agents, but require large trajectory corpora, executable environments, and costly feedback. In this paper, we propose STRATA, a framework that converts past terminal trajectories into stratified external knowledge and uses it to enhance agents without updating model parameters. STRATA organizes external knowledge into Skills that capture successful execution patterns and Memories that distill reusable failure lessons from contrasts between successful and failed trajectories. Given a new task, STRATA adopts an agentic retrieval strategy to select relevant Skills and Memories from the library, then assembles them into the terminal agent prompt while keeping the backbone model unchanged. This design allows the agent to use both positive execution guidance and negative failure evidence when planning commands. Experiments on Terminal-Bench 1.0 and Terminal-Bench 2.0 show improvements across both weaker and stronger backbone models. On Terminal-Bench 2.0, STRATA achieves a mean pass rate of 63.3% across three runs with DeepSeek-V4-Pro, outperforming Agent Workflow Memory by 9.4 percentage points under a matched evaluation protocol. These results suggest that structured trajectory-derived knowledge is an effective way to improve terminal agents on long-horizon execution tasks without model retraining.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Terminal Agent, Agent Skill, Agent Memory

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_cfNBOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

STRATA turns past terminal-task trajectories into knowledge and adds it to the prompt of a new task, without changing the model weights. Successful trajectories become multi-step procedures called Skills. Successful and failed trajectories from the same task are compared to produce Memories, which state when a rule applies and how to correct a mistake. Both kinds of knowledge are tagged with one of four capability layers, from command syntax to long-horizon planning. Before each new task, an auxiliary model picks the relevant entries. On Terminal-Bench 2.0, the paper compares STRATA with three experience-based methods (AWM, ExpeL, and ReasoningBank) and uses ablations to test the role of knowledge content, entry selection, and layer tags.

Critical strengths

Strength 1: STRATA gives a clear gain under a shared evaluation setup

In Table 2, STRATA raises the Pass Rate of DeepSeek-V4-Pro from 53.2% to 63.3%, above the best baseline, ReasoningBank (59.2%). The gain appears on both Medium and Hard tasks. The same library, built with DeepSeek-V4-Pro, also raises Qwen3-32B from 4.1% to 10.9% without fine-tuning. All methods share the same trajectories, the same model for building knowledge, the same scaffold and verifier, and the same prompt budget, so they are compared on equal terms.

Strength 2: The ablations show that how knowledge is organized and selected matters

In Table 3, keeping the knowledge but removing the layer tags lowers the Pass Rate to 57.3%. Keeping the library but switching to embedding retrieval lowers it to 50.6%. Generic CLI tips with similar format, topics, and length do not give the same gain. So the gain is not just from a longer prompt.

Critical weaknesses

Weakness 1: Paired extraction and distillation are not tested on their own

The paper never compares paired extraction with extracting lessons from successful and failed trajectories separately while keeping everything else fixed. The ReasoningBank comparison changes both how knowledge is extracted and how it is retrieved, so the two effects cannot be separated. The raw-trajectory ablation in Section 5.3 skips distillation, but it also picks trajectories by similarity instead of by an auxiliary model and leaves out failed trajectories. So its 4.9-point gap is not due to distillation alone. Memory-only reaches 61.8%, close to the full method's 63.3%, but this does not show how much the pairing step adds.

Weakness 2: The new contribution over AutoGuide is not clear

AutoGuide also compares trajectories with different outcomes on the same task, writes guidelines with conditions for when they apply, and picks rules based on the current execution history. STRATA cites AutoGuide but does not say which of its designs AutoGuide already has and which are new. The layer ablation already shows that capability layers help. The authors could start from this result and state which designs are new, such as the layers and the split into Skills and Memories.

Questions for authors
  1. (Related to Weakness 1) Have you compared paired and non-paired extraction with everything else fixed? Section 5.3 credits the 4.9-point gap over raw trajectories to distillation, but that comparison also changes entry selection and leaves out failed trajectories. What evidence shows that the gap comes from distillation?

  2. (Related to Weakness 2) Compared with AutoGuide, which new designs in STRATA matter most? Which existing results directly show that these designs help?

3 comments

Thanks for the insightful feedback! We will respond to your questions point by point below.

1. Effectiveness of success–failure paired memory distillation

Our additional experiments support the effectiveness of success–failure paired memory distillation. With STRATA's retrieval procedure, memory presentation and context budget unchanged, Paired achieves 63.30% average success versus 57.68% for Independent.

Both conditions draw from the same source trajectory pool and use a full-batch size of 24 trajectories. Each Paired batch contains 12 successful–failed pairs from the same tasks. Each full Independent batch contains either 24 successful trajectories or 24 failed trajectories. Thus, the two outcomes of a task never appear together in an Independent extraction input. The remaining retrieval, presentation and context-budget settings follow STRATA. Submitted paper: Sections 3.3–3.4, 4.4–4.5 and 5.1. After duplicate consolidation and support-frequency filtering, Independent retains 313 Memories, compared with 168 for Paired. Both conditions use their complete resulting memory pools. Table R1 reports the counts at each level, together with the three-run results.

Table R1. Memory extraction comparison. The layer columns give memory entry counts. SD is the sample standard deviation across three runs of 89 tasks.

ExtractionL1L2L3L4TotalSuccesses per run /89Success rate ± SD
Paired1972344316855, 54, 6063.30% ± 3.61 pp
Independent51132775331350, 51, 5357.68% ± 1.72 pp

LLM extraction and subsequent consolidation do not guarantee identical entry counts at every capability level. We therefore retain the naturally produced libraries without imposing layer quotas. The libraries have similar broad layer distributions, indicating similar coverage of capability levels. One possible explanation for the larger Independent library is that it summarizes successful practices and failure risks separately. Related experience can consequently survive consolidation as rules with different wording, types or contexts, potentially retaining more redundant formulations. Paired presents the two outcomes together and links the failed condition to a successful alternative in one rule, helping the extractor focus on the distinction that matters for the task.

Independent contains 86.3% more entries, yet its average success rate is 5.62 percentage points lower. Under the same context-budget specification, a larger library does not automatically provide more useful guidance to the agent. Same-task success–failure comparison gives Paired a direct reference for identifying the conditions and actions that distinguish the outcomes. The observed combination of a smaller library and higher average success is consistent with this advantage of paired extraction.

Overall, the additional experiments support the effectiveness of STRATA's success–failure paired memory distillation. Compared with independent extraction, it produces a smaller memory library while achieving a 5.62-point higher average success rate.

2. Distinctive designs relative to AutoGuide and their empirical support

The two most important designs that distinguish STRATA from AutoGuide are its shared capability layers and its explicit treatment of procedures and contrastive lessons as complementary knowledge resources. Our completed repeated ablations support the value of the layers and of providing both knowledge types. Catalog-based selection supports their use for a new task.

AutoGuide already extracts conditional guidelines from contrasting same-task trajectories and selects guidance using the current execution history. It also learns from offline experience for transfer to non-overlapping test tasks. These are shared foundations. The distinctions concern the representation, organization and use of that knowledge, as summarized below. AutoGuide: Sections 3.1–3.3 and 4.1.3.

Table R2. Method differences relevant to the reviewer's question.

DesignAutoGuideSTRATA
Knowledge representationContext-conditioned guidelines derived from trajectory contrastsSuccess-derived Skills encode action sequences and their purposes. Memories explicitly retain the failed action, successful alternative and applicability context. Sections 3.4 and 4.3–4.4.
Knowledge organizationA dictionary indexed by context, with matching of equivalent contextsBoth resources share L1–L4 capability tags, alongside a separate topic index. Tags enter catalog selection and prompt assembly. Section 3.3, Table 1 and Section 4.5.
Knowledge selectionContext matching and LLM guideline selection during executionBefore execution, an LLM selects categories, browses catalog summaries and reads candidate entries, separately for Skills and Memories. Section 4.5.

Capability layers specify the decision level at which knowledge applies, from command syntax to long-horizon planning. Removing these tags lowers the three-run mean success rate by 2.62 percentage points, supporting their value beyond topic labels. Submitted paper: Section 3.3, Table 1 and Section 5.3, Table 3. Skills provide procedures distilled from successful trajectories, while Memories capture applicability conditions and corrective lessons from success–failure contrasts. Removing Memories or Skills lowers the three-run mean by 3.37 and 2.25 points, respectively, supporting their complementary value. Submitted paper: Section 3.4, Figure 2 and Sections 4.3–4.5. Before execution, catalog-based selection checks entries' roles, capability levels and content. Replacing it with embedding top-k over the same library lowers the mean by 9.36 points. This supports the procedure within STRATA. Submitted paper: Section 4.5 and Section 5.3, Table 3.

Table R3. Evidence for STRATA's organization, knowledge resources and selection. Success rate is the three-run mean ± sample SD on 89 tasks.

ConfigurationDesign testedSuccess rate ± SDΔ, pp
Full STRATAComplete configuration63.30% ± 3.61 pp—
FlatRemove capability-layer tags60.67% ± 4.90 pp−2.62
Skill-onlyRemove Memory Pool59.93% ± 1.72 pp−3.37
Memory-onlyRemove Skill Library61.05% ± 3.43 pp−2.25
Embedding top-kReplace catalog-based selection53.93% ± 2.97 pp−9.36

Overall, STRATA contributes a capability-layered organization of complementary procedures and contrastive lessons, together with a selector that reads this structure. The repeated ablations provide empirical support for retaining the layers, supplying both knowledge types and using catalog-based selection. These are the concrete designs and corresponding results that distinguish our contribution beyond the shared idea of learning conditional guidance from trajectory contrasts.

Thank you again for your feedback. We hope our responses have clearly addressed your questions and concerns.

Thank you for the detailed response and the additional experiments. The controlled comparison between paired and independent extraction in Table R1 clearly shows that pairing same-task outcomes yields a more compact and effective memory pool. The design comparison with AutoGuide in Table R2 and the repeated ablations in Table R3 also clarify the roles of capability layers, complementary knowledge types, and catalog-based selection. These results address my main questions, and I will keep my positive score.

Reviewer_8DgfOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

STRATA augments frozen terminal agents with reusable Skills from successful trajectories and Memories from success–failure pairs on the same task. Both are organized by capability level and selected through an LLM-based retriever before execution. On Terminal-Bench 2.0, it improves DeepSeek-V4-Pro from 53.2% to 63.3% and Qwen3-32B from 4.1% to 10.9%.

Critical strengths

The improvement over the unaugmented agent is substantial. Sharing the trajectory pool, construction model, scaffold, and knowledge budget across methods makes the main comparisons useful.

The paper includes relevant ablations covering content, retrieval, layer tags, and the two knowledge types. The appendix also documents baseline adaptations, individual runs, costs, and detailed source–benchmark overlap checks.

The frozen library transfers to a second backbone, providing evidence that the extracted knowledge is useful beyond the model used to construct it.

Critical weaknesses

The experiments do not isolate the benefit of the central contrastive-memory design. The comparison with ReasoningBank changes both memory construction and retrieval, while removing the Memory Pool removes failure-derived knowledge altogether. Neither establishes whether pairing successful and failed attempts produces better knowledge than extracting lessons independently from the same trajectories. This matters because same-task contrastive learning already appears in ExpeL, AutoGuide, and ReasoningBank’s parallel MaTTS. A matched extraction comparison using the same retriever would help establish STRATA’s methodological contribution; the current results establish the effectiveness of the complete system more clearly than its distinctive design.

The evidence for stratification and complementary knowledge types is too uncertain to support the main conclusions. Appendix A.2 confirms that the ablations are single runs, compared against a three-run full-system mean. Memory-only achieves 61.8%, versus 63.3% ± 3.6% for full STRATA, leaving the added value of Skills unresolved. Removing layer tags produces a larger drop, but that result also lacks replication. Because stratification and the combination of Skills and Memories are central contributions, uncertainty about their benefits materially weakens the case for acceptance. Repeated runs of these targeted ablations would be appreciated.

Questions for authors

Does paired extraction outperform independent extraction from the same trajectory pool when retrieval, formatting, and context budget are held fixed?

Do the gains over Memory-only and the flat library persist across repeated runs, and which tasks consistently benefit?

If ReasoningBank-style memories use STRATA’s catalog selector, how much of the performance gap remains?

4 comments

Thanks for the insightful feedback! We will respond to your questions point by point below.

1. Paired versus independent memory extraction

Keeping STRATA's retrieval procedure, memory presentation format and context budget unchanged, we have completed an Independent extraction comparison. These results demonstrate an advantage of success–failure paired extraction in this comparison: a smaller memory library and a higher average task success rate.

Both conditions draw from the same source trajectory pool and use a full-batch size of 24 trajectories. Each Paired batch contains 12 successful–failed pairs from the same tasks. Each full Independent batch contains either 24 successful trajectories or 24 failed trajectories. Thus, the two outcomes of a task never appear together in an Independent extraction input. The remaining retrieval, presentation and context-budget settings follow STRATA. *Submitted paper: Sections 3.3–3.4, 4.4–4.5 and 5.1.*After duplicate consolidation and support-frequency filtering, Independent retains 313 Memories, compared with 168 for Paired. Both conditions use their complete resulting memory pools. Table R1 reports the counts at each level, together with the three-run results.

Table R1. Memory extraction comparison. The layer columns give memory entry counts. SD is the sample standard deviation across three runs of 89 tasks.

ExtractionL1L2L3L4TotalSuccesses per run /89Success rate ± SD
Paired1972344316855, 54, 6063.30% ± 3.61 pp
Independent51132775331350, 51, 5357.68% ± 1.72 pp

LLM extraction and subsequent consolidation do not guarantee identical entry counts at every capability level. We therefore retain the naturally produced libraries without imposing layer quotas. The libraries have similar broad layer distributions, indicating similar coverage of capability levels.One possible explanation for the larger Independent library is that it summarizes successful practices and failure risks separately. Related experience can consequently survive consolidation as rules with different wording, types or contexts, potentially retaining more redundant formulations. Paired presents the two outcomes together and links the failed condition to a successful alternative in one rule, helping the extractor focus on the distinction that matters for the task.

Independent contains 86.3% more entries, yet its average success rate is 5.62 percentage points lower. Under the same context-budget specification, a larger library does not automatically provide more useful guidance to the agent. Additional entries may overlap in content or leave the conditions for applying a rule unclear. Same-task success–failure comparison gives Paired a direct reference for identifying the conditions and actions that distinguish the outcomes. The observed combination of a smaller library and higher average success is consistent with this advantage of paired extraction.

Overall, Paired uses 168 Memories to achieve 63.30% average success, compared with 57.68% for Independent using 313 Memories. These results show an advantage for the Paired configuration in both memory-library size and average task success.

2. Repeated ablations of stratification and knowledge types

We have completed the repeated ablations, and Full STRATA retains the highest average success rate. Computational constraints limited the original ablations to one run, as disclosed in Appendix A.2. We have now added two runs for each of the six ablations in Table 3, bringing all six to three runs. Table R2 also includes the three-run paired-extraction ablation from our first response.

Table R2. Complete three-run ablations, following the structure of the submitted Table 3.

AblationChangeEasy (4)Med. (55)Hard (30)Avg. ± SD (%)Δ (pp)
Zero-shotw/o knowledge91.7 ± 14.463.0 ± 2.130.0 ± 5.853.18 ± 2.34-10.11
Contentw/o trajectory content91.7 ± 14.467.3 ± 3.630.0 ± 5.855.81 ± 3.61-7.49
w/o distillation83.3 ± 14.468.5 ± 2.831.1 ± 3.856.55 ± 1.72-6.74
w/o paired extraction91.7 ± 14.470.3 ± 2.130.0 ± 0.057.68 ± 1.72-5.62
Selectionw/o agentic retrieval91.7 ± 14.467.3 ± 6.624.4 ± 5.153.93 ± 2.97-9.36
Organizationw/o layer tags100.0 ± 0.072.7 ± 6.633.3 ± 3.360.67 ± 4.90-2.62
Componentsw/o Memory Pool83.3 ± 14.472.7 ± 1.833.3 ± 3.359.93 ± 1.72-3.37
w/o Skill Library91.7 ± 14.473.9 ± 5.633.3 ± 3.361.05 ± 3.43-2.25
Full STRATA—91.7 ± 14.475.2 ± 3.837.8 ± 5.163.30 ± 3.61—

The three-run averages show that Full's advantages over Flat and Memory-only are concentrated in medium and hard tasks. On medium tasks, Full achieves 75.2%, compared with 72.7% for Flat and 73.9% for Memory-only. On hard tasks, Full achieves 37.8%, compared with 33.3% for both. On easy tasks, Full and Memory-only both achieve 91.7%, while Flat achieves 100.0%. In terms of task content, the case study in Appendix A.7, Table 14 illustrates the value of STRATA for terminal workflows involving multi-step processing, strict output formats and final artifact verification. In the incident-response example, Skills organize file validation and report generation, while Memories guide error reporting and output-field selection, reducing corrective work during execution.

Overall, the three-run comparisons retain Full's average advantages of 2.62, 3.37 and 2.25 percentage points over Flat, Skill-only and Memory-only. The repeated results support the benefit of combining both knowledge types and organizing them by capability level.

3. Combining ReasoningBank with STRATA's catalog selector

We evaluated the combination of ReasoningBank (RB) with an adapted STRATA catalog selector, alongside embedding top-k retrieval for both knowledge libraries. Full STRATA remains the best-performing configuration, achieving 63.30% average success.

For offline RB construction, we follow the original method's success- and failure-specific extraction prompts, the limit of three memory items per trajectory, and the title–description–content format. The extracted entries are accumulated across source trajectories. Our bank contains 14,577 retrieval entries from 4,879 trajectory records and is frozen before evaluation. Its size therefore reflects RB's per-trajectory extraction and accumulation procedure. See ReasoningBank Section 3.2 and Appendix A.1–A.2.

RB does not provide STRATA's category and capability-level catalog. We retain RB's extracted memory content and use a broad embedding-retrieved candidate set to make LLM selection practical at this scale. Qwen3-Embedding-8B retrieves K = 50 candidates, ten times the maximum final selection. DeepSeek-Flash then reads their full text and selects at most five for the execution agent, within the same 7,200-character injection budget. This is our adaptation of candidate generation to RB's existing knowledge organization.

Table R3. Knowledge library and selection method. Each cell reports the three-run mean success rate ± sample SD.

Knowledge libraryEmbedding top-k retrievalLLM-based selection
STRATA53.93% ± 2.97 pp63.30% ± 3.61 pp
ReasoningBank59.18% ± 3.43 pp58.80% ± 1.30 pp

RB+LLM and RB with embedding retrieval have similar average performance. Their means are 58.80% and 59.18%, respectively. The 0.37-point difference is small relative to their run-to-run SDs of 1.30 and 3.43 points. Additional LLM selection can only act on the entries in its candidate set. Its benefit therefore depends both on what knowledge was retained during construction and on whether candidate retrieval makes the relevant entries available. Even when useful knowledge is present in the bank, it cannot be selected if candidate retrieval omits it. The observed results show that adding LLM selection through this adaptation does not reproduce Full STRATA's performance. STRATA's catalog selection achieves a 9.36-point higher mean than embedding top-k. This supports checking whether an entry applies to the task in addition to its semantic similarity. Submitted paper: Section 4.5.

Overall, Full STRATA retains the highest mean, exceeding RB with embedding retrieval by 4.12 points and RB+LLM by 4.49 points. Together with the STRATA top-k ablation, these results support the value of integrating knowledge construction, capability-based organization and task-specific selection. In the tested combination, retaining RB's original memories and adding LLM selection does not close the gap to the complete STRATA method.

Thank you again for your feedback. We hope our responses have clearly addressed your questions and concerns.

Thank you again for your thoughtful review. We would like to follow up on our response and would greatly appreciate your thoughts on whether it addresses your main concerns. If any points remain unclear or unresolved, we would be happy to provide further clarification. If you find that our response has sufficiently addressed your concerns, we would be grateful if you could reconsider your rating in light of these clarifications.

Thank you for your time and consideration.

Reviewer_r3mPOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

STRATA is a training-free framework that turns past terminal trajectories into stratified external knowledge for a frozen agent. Successful trajectories become Skills, reusable multi-step procedures, while contrasts between successful and failed attempts on the same task become Memories about when a procedure fails and what succeeded instead. Both are organized by capability level, from command syntax to long-horizon planning, and an agentic retriever selects relevant entries for each task.

Critical strengths
  • The paper addresses an important problem in terminal agents: successful procedures alone may not explain when a plausible action fails or how to recover. Recording the failed action, successful alternative, and applicable context is a sensible memory design.

  • Clean experimental design which is repeated across various models, showing consistent improvements.

  • The matched static-tip control is useful. Generic CLI advice of similar format, topics, and length does not reproduce the improvement, suggesting that simply adding more instructions is insufficient.

Critical weaknesses
  • It is unclear whether the gains come from contrastive memory construction or better retrieval. Learning from successful and failed trajectories already exists in ExpeL, including comparisons between successful and failed attempts on the same task. STRATA adds separate resources, capability layers, contextual entries, and agentic retrieval. Table 3 shows the largest drop when replacing agentic retrieval, but removing the Memory Pool does not isolate the benefit of paired success–failure distillation. Authors should compare paired and independently extracted memories while keeping the retriever, layers, entry counts, and prompt budget fixed.

  • Section 5.4 attributes the remaining failures to execution rather than retrieval, but successful knowledge injection and library usage do not establish that the retrieved guidance was relevant or sufficient. How did the authors verify that retrieval was correct on failed tasks? Removing confounders can help understand effects of irrelevant retrieval, missing knowledge, incorrect guidance, and failure to follow guidance instruction.

Questions for authors
  • Does paired success–failure distillation outperform independent extraction under the same retrieval and prompt budgets?

  • How much of the advantage remains when ReasoningBank uses STRATA’s agentic retriever?

3 comments

Thanks for the insightful feedback! We will respond to your questions point by point below.

1. Paired extraction under the same retrieval and prompt budgets

With STRATA's retrieval procedure, capability layers, memory presentation and context budget held unchanged, the Paired configuration achieves 63.30% average success, compared with 57.68% for Independent. We have completed the requested independent-extraction comparison, and the Paired configuration has a 5.62-point average advantage.

Both conditions draw from the same source trajectory pool and use a full-batch size of 24 trajectories. Each Paired batch contains 12 successful–failed pairs from the same tasks. Each full Independent batch contains either 24 successful trajectories or 24 failed trajectories. Thus, the two outcomes of a task never appear together in an Independent extraction input. The remaining retrieval, presentation and context-budget settings follow STRATA. Submitted paper: Sections 3.3–3.4, 4.4–4.5 and 5.1. After duplicate consolidation and support-frequency filtering, Independent retains 313 Memories, compared with 168 for Paired. Both conditions use their complete resulting memory pools. Table R1 reports the counts at each level, together with the three-run results.

Table R1. Memory extraction comparison. The layer columns give memory entry counts. SD is the sample standard deviation across three runs of 89 tasks.

ExtractionL1L2L3L4TotalSuccesses per run /89Success rate ± SD
Paired1972344316855, 54, 6063.30% ± 3.61 pp
Independent51132775331350, 51, 5357.68% ± 1.72 pp

LLM extraction and subsequent consolidation do not guarantee identical entry counts at every capability level. We therefore retain the naturally produced libraries without imposing layer quotas. The libraries have similar broad layer distributions, indicating similar coverage of capability levels. One possible explanation for the larger Independent library is that it summarizes successful practices and failure risks separately. Related experience can consequently survive consolidation as rules with different wording, types or contexts, potentially retaining more redundant formulations. Paired presents the two outcomes together and links the failed condition to a successful alternative in one rule, helping the extractor focus on the distinction that matters for the task.

Independent contains 86.3% more entries, yet its average success rate is 5.62 percentage points lower. Under the same context-budget specification, a larger library does not automatically provide more useful guidance to the agent. Additional entries may overlap in content or leave the conditions for applying a rule unclear. Same-task success–failure comparison gives Paired a direct reference for identifying the conditions and actions that distinguish the outcomes. The observed combination of a smaller library and higher average success is consistent with this advantage of paired extraction.

Overall, Paired uses 168 Memories to achieve 63.30% average success, compared with 57.68% for Independent using 313 Memories. These results show an advantage for the Paired configuration in both memory-library size and average task success.

2. Remaining gap with an adapted STRATA agentic retriever

A 4.49-percentage-point gap remains when the ReasoningBank (RB) library uses our adapted STRATA agentic selector. We evaluated this combination alongside embedding top-k retrieval for both libraries, and Full STRATA retains the highest mean.

For offline RB construction, we follow the original method's success- and failure-specific extraction prompts, the limit of three memory items per trajectory, and the title–description–content format. The extracted entries are accumulated across source trajectories. Our bank contains 14,577 retrieval entries from 4,879 trajectory records and is frozen before evaluation. Its size therefore reflects RB's per-trajectory extraction and accumulation procedure. See ReasoningBank Section 3.2 and Appendix A.1–A.2.

RB does not provide STRATA's category and capability-level catalog. We retain RB's extracted memory content and use a broad embedding-retrieved candidate set to make LLM selection practical at this scale. Qwen3-Embedding-8B retrieves K = 50 candidates, ten times the maximum final selection. DeepSeek-Flash then reads their full text and selects at most five for the execution agent, within the same 7,200-character injection budget. This is our adaptation of candidate generation to RB's existing knowledge organization.

Table R2. Knowledge library and selection method. Each cell reports the three-run mean success rate ± sample SD.

Knowledge libraryEmbedding top-k retrievalLLM-based selection
STRATA53.93% ± 2.97 pp63.30% ± 3.61 pp
ReasoningBank59.18% ± 3.43 pp58.80% ± 1.30 pp

RB+LLM and RB with embedding retrieval have similar average performance. Their means are 58.80% and 59.18%, respectively. The 0.37-point difference is small relative to their run-to-run SDs of 1.30 and 3.43 points. Additional LLM selection can only act on the entries in its candidate set. Its benefit therefore depends both on what knowledge was retained during construction and on whether candidate retrieval makes the relevant entries available. Even when useful knowledge is present in the bank, it cannot be selected if candidate retrieval omits it. The observed results show that adding LLM selection through this adaptation does not reproduce Full STRATA's performance. STRATA's catalog selection achieves a 9.36-point higher mean than embedding top-k. This supports checking whether an entry applies to the task in addition to its semantic similarity. Submitted paper: Section 4.5.

Overall, Full STRATA retains the highest mean, exceeding RB with embedding retrieval by 4.12 points and RB+LLM by 4.49 points. Together with the STRATA top-k ablation, these results support the value of integrating knowledge construction, capability-based organization and task-specific selection.

Thank you again for your feedback. We hope our responses have clearly addressed your questions and concerns.

Thank you again for your thoughtful review. We would like to follow up on our response and would greatly appreciate your thoughts on whether it addresses your main concerns. If any points remain unclear or unresolved, we would be happy to provide further clarification.

Thank you for your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.