Core Concepts
The author proposes a novel memory scheme, CREEM, that blends past memories and refines outdated information to improve chatbot responses by ensuring an evolving long-term memory.
Abstract
The content discusses the importance of constructing a long-term memory for chatbots that can adapt to changes in speakers over time. It introduces the CREEM model, which blends past memories and refines outdated information to create a more informed and coherent long-term memory. The paper evaluates the performance of CREEM against existing methods using specific criteria and datasets, showcasing its superiority in integrating, maintaining consistency, achieving sophistication, and ensuring conciseness in memory construction.
Stats
CC DATASET: SumMemMSC - 3.16; REBOTCC - 4.01; CREEM - 4.05
MSC DATASET: SumMemMSC - 3.44; REBOTCC - 3.86; CREEM - 4.02
Quotes
"We propose a novel memory scheme for long-term conversation, CREEM."
"Our goal is to make an ever-evolving memory that effectively integrates long-term memory."
"CREEM continuously refines its memory every time there is new information to remember."