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

Problem

Language agents currently exhibit limitations in interactive environments, particularly in tasks that require long-horizon state tracking and the ability to recover from failures. This paper addresses these gaps by introducing cognitive extensions that enhance the operational capabilities of language agents in such contexts. The work is presented as a preprint and has not undergone peer review.

Method

The authors propose an architecture named SwiftSage, which integrates two key components: the Adaptive Memory Module (AMM) and the Self-Reflection Module (SRM). The AMM facilitates salience-gated episodic storage and trigger-driven retrieval, allowing the agent to manage and utilize past experiences effectively. The SRM is designed for bounded execution-time validation and corrective intervention, enabling the agent to reflect on its actions and adjust accordingly. The system was evaluated using data sourced from ScienceWorld, and various configurations were tested: a baseline model, a model with AMM, a model with SRM, and the full system combining both modules.

Results

The results demonstrate significant improvements over the baseline model: the mean final score achieved was 64.62, with a success rate of 43.17% and a successful-step efficiency of 19.33 steps. Notably, the SRM was identified as the strongest contributor to these performance enhancements when evaluated as a standalone component.

Limitations

The authors do not report any limitations in their study, and no obvious limitations are identified in the available text.

Why it matters

The introduction of cognitive extensions such as AMM and SRM in language agents has important implications for the development of more robust interactive systems. By improving long-horizon state tracking and failure recovery, this work paves the way for more sophisticated applications of language agents in real-world scenarios, potentially enhancing user experience and operational efficiency.