- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
The paper proposes SAGA, a memory framework for LLM agents. Trajectories are abstracted into four levels: raw trajectories, episodic summaries, procedures, and principles with applicability conditions. At test time, retrieved principles are rewritten into task-specific constraints, and some principles are also used to check candidate actions and trigger resampling. The method is evaluated on ScienceWorld and ALFWorld, both with frozen API models and with SFT + RL on Qwen2.5-7B.
Critical strengths
The observation that injecting abstract principles directly can hurt, while instantiating them for the current task helps, is interesting and useful for work on agent memory beyond this paper.
The memory construction is careful, and the appendix is honest about several confounds and limitations.
Critical weaknesses
-
The evidence does not support the main claims. In the ablation, most of the gain comes from simply retrieving raw trajectories; the parts that are new in this paper add only a few points on top of that, from single runs without variance estimates. The generalization claim is also not supported: on ALFWorld the strong model already solves the unseen task types without memory, and the 7B improvement cited in the main text actually comes from the trajectory-only control, as the appendix acknowledges.
-
The gains cannot be clearly attributed to the method. The training-based results use trajectories from much stronger models for SFT, with no control trained on the same data without the knowledge context, so they may largely reflect distillation. Baselines are copied from other papers, while the authors' own harness gives a noticeably different score for the same base model. The closest prior work (CLIN, AutoGuide, AutoManual) is discussed but not compared.
Questions for authors
- Can you report results over multiple seeds, or with confidence intervals, for the main and ablation tables?
- How was the reported round or checkpoint chosen? Please also report last-round results.
- Can you add SFT on the same trajectories without knowledge context, and a setting where the same model builds and uses the memory?
4 comments
Response to Reviewer Zxva (Part II)
W2(a) & Q3: Separating distillation, knowledge conditioning, and builder strength
The training-based results contain a substantial teacher-action distillation component. On ScienceWorld, the SFT targets are reconstructed from Qwen3.8-Max evolution trajectories using the score-dependent trajectory-prefix rules specified in the Appendix. To separate teacher-action distillation from hierarchical knowledge conditioning, all SFT variants use the same selected teacher trajectories and action targets. They also share the Qwen2.5-7B actor, GRPO task pool, reward, rollout budget, and evaluation setting. GRPO starts from each SFT checkpoint and uses new on-policy rollouts rather than replaying the teacher trajectories.
| Qwen2.5-7B configuration | Teacher actions | Knowledge builder | SFT | SFT + GRPO |
|---|---|---|---|---|
| Base model | No | None | 9.6 | — |
| Same teacher actions, no knowledge context | Yes | None | 53.8 | 62.7 |
| Same-model SAGA knowledge | Yes | Qwen2.5-7B | 55.1 | 63.8 |
| Full SAGA knowledge | Yes | Qwen3.8-Max | 58.1 | 66.9 |
The increase from 9.6 to 53.8 shows the large contribution of teacher-action distillation. With teacher actions fixed, knowledge constructed by Qwen2.5-7B further improves SFT by 1.3 points and the final policy by 1.1 points. Replacing the same-model builder with Qwen3.8-Max adds another 3.0 points after SFT and 3.1 points after GRPO. In total, full hierarchical knowledge conditioning improves the matched no-knowledge control by 4.3 points after SFT and 4.2 points after GRPO.
These controls separate three sources of improvement: teacher-action distillation provides the main initial gain, hierarchical knowledge conditioning adds a further gain under fixed supervision, and the stronger builder improves the quality of that knowledge. The positive same-model control also shows that the knowledge pipeline remains useful without a stronger external builder. Revised Table 2 reports the complete attribution study.
W2(b): Same-harness comparison with prior methods
We thank the reviewer for identifying the ambiguity in the submitted comparison. The submission placed results copied from prior papers beside results produced by our harness, although their prompts, model wrappers, action parsers, backbones, and evaluation protocols differed. The revised manuscript separates reported cross-paper results from matched reimplementations.
Within each benchmark, all matched methods use the same actor, source experiences, task split, interaction budget, context budget, and evaluation harness. ScienceWorld uses Qwen3.8-Max. ALFWorld uses Qwen2.5-7B as the actor and Qwen3.8-Max for memory or rule construction. For AutoGuide, we follow its original success–failure trajectory pairing, context identification, guideline extraction, and top-2 guideline selection. Look and Pick2 remain excluded from the shared experience pool.
| ScienceWorld method | Avg. Score |
|---|---|
| No memory | 48.4 ± 0.5 |
| Trajectory retrieval | 63.7 ± 0.6 |
| ExpeL | 65.2 ± 0.9 |
| Memento | 66.4 ± 1.0 |
| ReasoningBank | 66.7 ± 0.8 |
| CLIN | 66.9 ± 0.9 |
| Submitted SAGA | 67.0 ± 0.7 |
| Revised SAGA | 72.4 ± 1.1 |
| ALFWorld method | Overall success rate |
|---|---|
| No memory | 45.5 |
| Trajectory-only | 57.5 |
| AutoGuide | 61.9 |
| AutoManual | 64.1 |
| Revised SAGA | 70.1 |
Submitted SAGA performs comparably to CLIN under the matched ScienceWorld setting. Revised SAGA exceeds CLIN by 5.5 points (95% CI: [4.1, 6.9], p = 0.002). On ALFWorld, revised SAGA exceeds AutoGuide by 8.2 points and AutoManual by 6.0 points. Revised Tables 3 and 6 report these matched comparisons, and the Appendix specifies the corresponding protocols.
Response to Reviewer Zxva (Part III)
Q2: Checkpoint selection and last-round performance
The reported 66.9% result was selected without using test-set performance. During GRPO, checkpoints were ranked only by their average reward on the training task pool. The checkpoint with the highest training reward was fixed before test evaluation and then evaluated once on the held-out test set. No separate validation split was introduced.
| Checkpoint | Selection rule | Test performance |
|---|---|---|
| Training-selected checkpoint | Highest average reward on the training task pool | 66.9% |
| Final checkpoint | Last GRPO checkpoint | 66.2% |
The final checkpoint remains within 0.7 points of the training-selected checkpoint. Figure 4 and the accompanying analysis now replace the ambiguous label “Best” with “Training-selected checkpoint,” state the selection rule, and report the final-checkpoint result.
Revision. Changes made specifically in response to Reviewer Zxva are highlighted in purple (#7B1FA2) in the revised manuscript. Revisions shared across multiple reviews are highlighted in orange (#B45309). The revision adds the five-memory ablation and statistical analysis in Table 7; the non-saturated held-out transfer study in Table 5; the matched distillation, knowledge-conditioning, and builder-strength controls in Table 2; the matched prior-method comparisons in Tables 3 and 6; and the checkpoint-selection rule and final-checkpoint result in Figure 4 and the accompanying text.
Follow-up on our response
Dear Reviewer Zxva,
Thank you again for your detailed assessment. We have addressed the concerns regarding variance and statistical reliability, non-saturated generalization, the attribution of distillation and knowledge conditioning, matched baseline comparisons, and checkpoint selection, and incorporated the corresponding results into the revised manuscript.
As the discussion period draws to a close, we would appreciate it if you could let us know whether our response and revisions have adequately addressed your concerns. If any issue remains unclear or unresolved, we would be grateful for the opportunity to provide further clarification.
Sincerely,
The Authors of Submission 814
Response to Reviewer Zxva (Part I)
We thank the reviewer for the detailed assessment and for identifying the central attribution and evaluation questions. We address the weaknesses (W) and questions (Q) in order below.
W1(a) & Q1: Gains beyond trajectory retrieval and statistical reliability
Raw trajectory retrieval provides a substantial initial improvement. The key question is whether SAGA adds a reliable gain beyond this strong control. We therefore repeated the complete ScienceWorld ablation using five independently constructed memories. Within each memory, all configurations use the same test instances, actor, source experiences, context budget, and evaluation harness.
The submitted full configuration improves over trajectory retrieval by 3.3 points, with a positive improvement in all five independently constructed memories. Outcome-validated SAGA further increases the improvement to 8.7 points (95% CI: [7.5, 9.9], paired p = 0.004). It also exceeds the submitted full configuration by 5.4 points (95% CI: [4.3, 6.5], paired p = 0.006).
These results clarify the roles of the components. Raw trajectories provide strong task-specific guidance. Hierarchical abstraction, task-conditioned instantiation, and action regulation add a consistent gain, while outcome validation produces the largest additional improvement by using later execution outcomes to reinforce, narrow, or exclude principles. Revised Table 7 reports mean ± SD for every ablation row and replaces the original single-run interpretation with the five-memory statistical analysis.
W1(b): Generalization with a non-saturated actor
The Qwen3.8-Max ALFWorld setting is saturated: the no-memory actor already obtains 100% on both Look and Pick2. SAGA retains 100% on Look and obtains 94.1% on Pick2 because one game reaches the context limit. The transfer analysis therefore uses the fixed Qwen2.5-7B actor, whose no-memory performance remains far from saturation.
The 57.5% result in the submission corresponds to trajectory retrieval, while the submitted complete hierarchy reaches 59.4%. Under the same fixed actor, held-out task types, interaction budget, and evaluation harness, outcome-validated SAGA reaches 70.1%. Look and Pick2 remain excluded from experience collection and principle validation, and the memory is frozen before evaluation.
Compared with trajectory retrieval, revised SAGA improves Look by 27.8 points, Pick2 by 23.5 points, and overall performance by 12.6 points. Revised Tables 4 and 5 correct the attribution of the 57.5% result, add the complete 7B comparison, and ground the transfer conclusion in the non-saturated Qwen2.5-7B setting.