Memory for cognitive agents

In dialogic interactions between humans and cognitive agents - typically implemented using large language models (LLMs) - the stateless nature of these models prevents the agent from building up experience and establishing a common ground with the human partner. Statelessness here means that these models do not store memories of previous conversations; there is no awareness of prior interactions. This raises the question of how experiences, knowledge, and insights gained through ongoing interactions between individuals and cognitive agents in their physical and social environments can be persistently stored by the agent, retrieved in a contextually appropriate manner during subsequent interactions, and effectively utilized as memories. The Adaptive Control of Thought-Rational (ACT-R) cognitive architecture, with its symbolic and sub-symbolic capabilities for retrieving information from declarative memory, enables the realization of human-like memory functions. The LLM used for the agent’s conversational contributions generates keywords related to the current conversation content as well as the actual recollection phrase. This is stored in a chunk of declarative memory and the recollection can be recalled associatively via the keywords. The persistent storage of memory chunks - including the activations of individual chunks generated through base-level learning - enables the formation of long-term memory and learning through preferred memories with higher activation over extended periods, as well as the ability to temporarily shut down the agent, restart it later, and continue to utilize the memories. Such shared experiences, made possible by shared memories and common ground, foster a deeper interaction between humans and machines, in which humans feel more perceived, thereby enriching human-robot interaction with a sense of connection and togetherness.