A Prompt Chaining Framework for Long-Term Recall in LLM-Powered Intelligent Assistant

DOI

Intelligent Assistants (IAs) often struggle with maintaining context in extended conversations, limiting their long-term recall capabilities and personalization. This study introduces a prompt chaining framework that integrates Large Language Models (LLMs) into IAs to address these issues. Our approach encompasses: (1) a formative study (N=30) analyzing IA-user interactions and identifying key challenges in long-term memory and personalization, (2) development of an LLM-based system with a novel prompt chaining mechanism for improved context retention, adaptable to various LLM architectures, and (3) implementation of a multi-step reasoning process to enhance context-aware and personalized responses. We evaluated the framework through quantitative analysis on multiple datasets, tested with different LLM backends, and conducted ablation studies to assess component contributions. The results show improvements in long-term recall, context awareness, and personalization across all tested models and datasets. A human evaluation study (N=47) indicated better Sensibleness, Consistency, and Personalization in IA response. This approach enhances IA memory capabilities and can adapt to different LLM backends.