This study develops a prompt chaining framework to enhance personalization and long-term recall capabilities in intelligent assistants (IAs). In the first phase, we conducted an online survey (n=83) using Likert scales and open-ended questions to anal...
This study develops a prompt chaining framework to enhance personalization and long-term recall capabilities in intelligent assistants (IAs). In the first phase, we conducted an online survey (n=83) using Likert scales and open-ended questions to analyze user expectations and privacy concerns regarding personalization, identifying four distinct user groups with varying attitudes. We then performed a controlled user study where participants (n=12) evaluated six IAs across five personalization-focused tasks, revealing specific limitations in current IAs. Based on these findings, we developed a prompt chaining framework that integrates Large Language Models (LLMs) with memory management techniques. The framework consists of six components: Conversation Manager, Context Extractor, Memory Module, Prompt Generator, LLM Interface, and Response Generator. It uses adaptive prompt chains (3-5 steps) for reasoning and implements a dual-layer memory architecture (short-term: 10 turns, long-term: 5,000 entries) for context retention. We evaluated the framework using three datasets with different LLM backends, and conducted ablation studies to validate the importance of each component. A subsequent human evaluation study (n=47) confirmed overall improvements in sensibleness, consistency, and personalization of responses. The results demonstrate the effectiveness of combining structured prompt engineering with memory management in addressing the limitations of current IAs.