Cognitive Extensions for Dual-Process Language Agents: Memory and Self-Reflection in Interactive Environments

TL;DR

The paper enhances a dual-process language agent with two modular cognitive extensions: an Adaptive Memory Module for storing and retrieving relevant past experiences, and a Self-Reflection Module for validating actions during execution. Experiments on ScienceWorld show that the full system outperforms all variants, with self-reflection being the most impactful standalone component.

Why It Matters

Improving language agents' reliability in interactive environments is critical for real-world deployment. This work demonstrates that runtime control and memory management are key bottlenecks, offering practical insights for building more robust autonomous agents.

Key Findings

• The Self-Reflection Module significantly improves performance by enabling bounded-time validation and corrective actions during execution.
• The Adaptive Memory Module enhances long-horizon task success, but its benefits are most evident when execution stability is already ensured.
• The full system achieves the highest mean score, success rate, and step efficiency, indicating synergistic benefits from combining both cognitive extensions.