Using Memory Contents of a Cognitive Model for Prompt Augmentation of a Large Language Model

Large Language Models (LLMs) are becoming increasingly widespread thanks to their broad application possibilities and good performance. However, reliable use, for example in the provision of information, is hampered by the fact that the utterances of an LLM are occasionally inappropriate, untrue or fictitious. In addition, LLMs are limited in their ability to make human-like judgments and conclusions, especially over several steps in complex tasks. We propose an approach for augmenting the prompt used in an LLM by means of the human-like judgment and decision-making capabilities inherent in cognitive architectures for a desired deployment scenario. In particular, we access the memory contents of an ACT-R model in order to use the knowledge stored there to constrain the system prompt of the LLM in such a way that the language model can correctly reproduce facts that are otherwise unknown to it. We exemplify the use of such an approach in a Human-Robot Interaction (HRI) scenario with the social robot Pepper.