Cognitive Extensions for Dual-Process Language Agents: Memory and Self-Reflection in Interactive Environments
João Meneses dos Santos, Arlindo L. Oliveira
cs.AI · cs.LG · cs.MA
Quick verdict
Adds memory + bounded self-reflection to SwiftSage on ScienceWorld; SRM drives most gains (51.67→64.33), but no significance tests leave full-vs-SRM open.
Summary
The paper extends SwiftSage, a dual-process language agent, with two feature-flagged modules evaluated on ScienceWorld: an Adaptive Memory Module (AMM) that salience-gates episodic writes and triggers retrieval at Swift-failure (T1) and Sage-planning (T4) points, and a Self-Reflection Module (SRM) providing pre-execution Gate-1 action validation, post-step stagnation detection, and a budget-bounded Critic whose corrective actions enter Sage's FIFO buffer. Four configurations (baseline, +AMM, +SRM, full) run on a shared substrate (Flan-T5-large Swift; Qwen2.5-7B-Instruct-1M Sage/Critic via vLLM; Letta archival memory) over tasks 0-29, up to 271 episodes each. The full system scores best (64.62 final, 43.17% success, 19.33 Steps@Succ), but +SRM nearly matches it (64.33/41.33%), while AMM alone adds only 4.1% and hurts Long tasks (-6.6%). Mechanism logs attribute SRM's gains mainly to Gate-1 filtering (17-20 drops/episode) rather than Critic calls (~1.3/episode); System 2 calls fall from 24.6 to 14.1 per episode. Raising the Critic budget from three to six degrades score from 64.33 to 52.95. Limitations include a single environment, hand-crafted thresholds, no variance or significance reporting, a populated-memory protocol, and no stated code release.
Review complete
Assessment at a glance
5
risk flags identified
● high 1
● medium 3
● low 1
4 strengths · 4 weaknesses · 5 author questions
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- The full system (SwiftSage+AMM+SRM) achieves the best aggregate performance on ScienceWorld: 64.62 mean final score, 43.17% success, and 19.33 steps per successful episode.
- SRM is the strongest standalone contributor (+24.5% score, success 23.99%→41.33%), implying execution-time control is the dominant bottleneck in this setting.
- AMM alone is only modestly beneficial (+4.1%) and hurts Long tasks, but becomes coherent when composed with SRM because stabilized execution reduces unstable states needing memory intervention.
- Unbounded reflection harms performance: raising the Critic budget from three to six calls lowers score from 64.33 to 52.95 across all task-length groups.
- Gains come from a safer fast path rather than more deliberation, as System 2 calls fall from 24.6 to 14.1 per episode while Swift-executed steps rise to 62.01%.
Strengths
Interpretable ablation design backed by mechanism-level logging
Feature-flagged modules over one shared execution substrate, plus Tables 5-7, let the authors attribute gains to mechanisms: Gate-1 drops (17.5-20.0/episode) dominate over effective Critic calls (~1.3/episode), directly supporting the 'continuous filtering, not frequent replanning' interpretation in Section 5.
Boundedness claims are validated empirically, not just asserted
Table 8 shows the Critic budget 3→6 sensitivity degrades score from 64.33 to 52.95 in every oracle-length group; the Swift T1-injection ablation (64.62→63.33, worst on Long tasks) ties a specific retrieval trigger to outcomes.
Honest protocol framing and limitations
Efficiency results cohere with the central claim
Weaknesses
No statistical treatment of results
Gate-1 relies on the environment-provided valid-action set
AMM's contribution is weak and partly negative, complicating the narrative
Protocol ambiguities and no code release
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
- All 4 strengths & 4 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What does the paper 'Cognitive Extensions for Dual-Process Language Agents' propose?
It adds two modular extensions to SwiftSage, a 'fast/slow thinking' LLM agent: AMM, an episodic memory module that stores only salient moments (successes, near-misses, failures) and retrieves them at specific recovery/planning points, and SRM, a self-reflection module that validates every action before execution, detects stagnation, and calls a Critic model at most a few times per episode to inject corrective actions. Both are tested via controlled ablations on the ScienceWorld text environment.
What results does the system achieve on ScienceWorld?
Across four configurations, the full system (memory + self-reflection) reaches a 64.62 mean final score, 43.17% task success, and 19.33 steps per successful episode, versus 51.67, 23.99%, and 24.48 for the baseline. However, adding only the self-reflection module nearly matches this (64.33, 41.33%), while adding only memory yields 53.83 (26.94%), so execution-time control accounts for most of the improvement.
Which matters more for LLM agents: memory or self-reflection?
In this paper's ScienceWorld experiments, self-reflection (SRM) matters far more: most gains come from pre-execution filtering of invalid or stale actions (about 17-20 filtered actions per episode) rather than from reflective replanning. Episodic memory (AMM) helps only modestly on its own and even hurts the longest tasks, becoming useful mainly as supporting evidence once execution is already stabilized.
Is code available for this paper?
The paper text reviewed here does not state any code or data release. It provides a reproducibility appendix listing models (Flan-T5-large for the fast proposer, Qwen2.5-7B-Instruct-1M via vLLM for the planner and Critic, Letta for memory storage) and protocol details, but no repository link is given.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figure 1 (architecture) is available only as a caption; the visual layout and any details conveyed only in the figure could not be assessed.
- The evaluation protocol says 'up to 271 episodes per configuration', so exact per-configuration sample sizes cannot be verified from the text.
- Stagnation thresholds, Gate-1 repair rules, and prompt templates are not fully specified in the text, limiting verification of the control logic.
- All tables and numbers were taken from the extracted text as given; no external verification of the reported results was possible.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.