Procedural Memory — References

The distinction between knowing-that and knowing-how, and the agent systems that accumulate re-usable know-how at runtime.

The distinction

• Declarative and Nondeclarative Memory: Multiple Brain Systems Supporting Learning and Memory — Larry R. Squire, Journal of Cognitive Neuroscience, 1992. The separation between knowledge that can be stated and skill that is expressed through performance. Cited by name.
• The Architecture of Cognition — John R. Anderson, Harvard University Press, 1983. A cognitive architecture in which procedural knowledge is represented as condition-action rules distinct from the declarative store, and in which practice changes the rules. Cited by name.

Accumulating know-how at runtime

• Voyager: An Open-Ended Embodied Agent with Large Language Models — Guanzhi Wang, Yuqi Xie, Yunfan Jiang, Ajay Mandlekar, Chaowei Xiao, Yuke Zhu, Linxi Fan and Anima Anandkumar, arXiv preprint, 2023 (arXiv:2305.16291). An agent that writes successful behaviours into a growing library of re-usable skills and retrieves them for later tasks. Cited by name.
• Reflexion: Language Agents with Verbal Reinforcement Learning — Noah Shinn, Federico Cassano, Edward Berman, Ashwin Gopinath, Karthik Narasimhan and Shunyu Yao, NeurIPS, 2023. Retaining a written lesson from a failed attempt and conditioning later attempts on it. Cited by name.
• Cognitive Architectures for Language Agents — Theodore R. Sumers, Shunyu Yao, Karthik Narasimhan and Thomas L. Griffiths, Transactions on Machine Learning Research, 2024 (arXiv:2309.02427). Names a procedural store alongside episodic and semantic stores and discusses what writing to it safely requires. Cited by name.

Why a learned routine stops working

• Catastrophic Interference in Connectionist Networks: The Sequential Learning Problem — Michael McCloskey and Neal J. Cohen, in The Psychology of Learning and Motivation, Volume 24, Academic Press, 1989. What happens to previously learned behaviour when a system continues learning without protecting what it already had. Cited by name.

No verified reference is recorded for promotion and demotion thresholds on learned routines, for environment fingerprinting, or for storing failed approaches as retrievable items. These are engineering practice without published study.