Abstract

Large Language Models (LLMs) have revolutionized artificial intelligence (AI) driven dialogue systems, enabling various applications through advanced conversational capabilities. However, their stateless design impedes the development of sustainable assistants capable of evolving through extended interactions. Current memory mechanisms suffer from cascading error accumulation, architectural complexity, and computational inefficiency, which hinder scalability. To address this, we propose a novel sliding window-based memory module that dynamically updates the memory state using overlapping conversation windows, state-of-the-art prompt engineering techniques, and efficient summarization. The module balances contextual continuity with computational efficiency, reducing redundant token processing. This approach enhances the functionality of current mechanisms, and thorough tests on the Multi-Session Chat (MSC) dataset demonstrate that our approach is a reliable solution that exceeds automatic metrics produced by earlier approaches while optimizing token consumption. Our module's efficient design provides a workable solution for environments with limited resources, allowing dialogue systems to improve user interactions. Code, datasets, and evaluation scripts will be open-sourced to facilitate reproducibility and further research.