Cognitive Extensions for Dual-Process Language Agents: Memory and Self-Reflection Modules Extend SwiftSage

Forum topic · 小凯 · 2026-09-18

Summary

This paper extends SwiftSage, a dual-process language agent combining a fast action proposer with a slower planner, with two modular cognitive extensions: an Adaptive Memory Module (AMM) providing salience-gated episodic storage and trigger-driven retrieval, and a Self-Reflection Module (SRM) providing bounded execution-time validation and corrective intervention. Both modules are feature-flagged extensions over the same execution substrate, enabling controlled ablations on the ScienceWorld benchmark. Across four configurations—baseline, baseline+AMM, baseline+SRM, and the full system—the full system achieves the best mean final score (64.62), success rate (43.17%), and step efficiency (19.33 steps), with SRM being the strongest individual contributor. The results suggest that execution-time control is the dominant bottleneck in interactive environments, while episodic memory becomes most valuable once the runtime loop is stabilized. Authored by João Meneses dos Santos and Arlindo L. Oliveira, the paper (arXiv:2609.19128) was released on 2026-09-16.

Paper Overview

Field: Machine Learning Authors: João Meneses dos Santos, Arlindo L. Oliveira Published: 2026-09-16 arXiv: 2609.19128

Original Abstract

Language agents remain brittle in interactive environments, where success requires long-horizon state tracking, valid action execution, and recovery from failed steps. We extend SwiftSage, a dual-process agent that combines a fast action proposer with a slower planner, using two modular cognitive extensions: an Adaptive Memory Module (AMM) for salience-gated episodic storage and trigger-driven retrieval, and a Self-Reflection Module (SRM) for bounded execution-time validation and corrective intervention. Both modules are implemented as feature-flagged extensions over the same execution substrate, enabling controlled ablations on ScienceWorld. Across four configurations—baseline, baseline+AMM, baseline+SRM, and the full system—the full system achieves the best mean final score (64.62), success rate (43.17%), and step efficiency (19.33 steps), with SRM as the strongest individual contributor.

Key Findings

Dual-process architecture: SwiftSage pairs a fast action proposer with a slower planner; this work adds two pluggable cognitive extensions.
Adaptive Memory Module (AMM): performs salience-gated episodic storage with trigger-driven retrieval.
Self-Reflection Module (SRM): provides bounded execution-time validation and corrective intervention.
Controlled ablations: both modules are feature-flagged over the same execution substrate, evaluated on ScienceWorld in four configurations.
Results: the full system achieves the best mean final score (64.62), success rate (43.17%), and success step efficiency (19.33 steps); SRM is the strongest standalone contributor.
Implication: execution-time control is the dominant bottleneck in this setting, while episodic memory becomes most valuable once the runtime loop is stabilized.