Encoding and Retrieving Memories for Persona Grounded Dialogues

November 14, 2024

Conversational AI systems that leverage large language models (LLMs) have achieved remarkable progress in natural language capabilities. However, due to the lack of persistent long-term memory, these models struggle to sustain meaningful longitudinal relationships requiring knowledge across interactions. This thesis proposes MemoraGPT - a dual memory architecture to augment LLMs for consistent persona-grounded dialogues in human-robot interaction scenarios. The solution encompasses three key modules - a Conversational Memory Module to consolidate memory from past conversations, a Facts Memory Module to encode and user-specific details and events, and a Controller Module to orchestrate all the memory processes. The integrated system facilitates targeted retrieval and context-aware text generation, aimed at optimizing coherence while sustaining continuity across multiple sessions. Through human evaluation studies spanning repeated interactions, metrics capturing perceived sociability, competence and others, will assess the impact of the memory components on the quality of user experience. The goal is to uncover nuanced facets of human-like conversational capacities to enrich dialog agents for naturalistic and engaging human-robot relationships. With memory augmentation, this research contributes towards advancing the sophistication of interactions in the evolving landscape of conversational AI and social robotics.

Authors in the community:

• 

  D

  Daniel Alexandre Quintas Lopes

  ist193696

Supervisors of this institution:

• 

  J

  Joana Carvalho Filipe de Campos

  ist155848

electrical-engineering-electronic-engineering-information-engineering - Electrical engineering, electronic engineering, information engineering

eng - English

Instituto Superior Técnico