Large Language Models currently face inherent limitations in managing context despite advances that extend context windows to millions of tokens. Traditional Retrieval-Augmented Generation (RAG) systems rely on passive retrieval and lack the dynamic, task-driven memory management seen in human cognition. The Cognitive Workspace paradigm draws on cognitive science theories such as Baddeley’s working memory model and Clark’s extended mind thesis to propose a system that actively manages memory through deliberate information curation, hierarchical cognitive buffers, and task-driven context optimization. This approach enables persistent working states and adapts dynamically to cognitive demands, addressing the shortcomings of existing techniques like Infini-attention and StreamingLLM, which, while extending context length, do not incorporate metacognitive awareness or active planning. Empirical studies demonstrate that Cognitive Workspace achieves an average memory reuse rate of 58.6%, a stark contrast to the 0% reuse in traditional RAG, and yields a 17-18% net efficiency gain despite increased operation counts. Statistical analyses confirm the robustness of these improvements with highly significant p-values and large effect sizes. The framework synthesizes insights from over 50 recent studies, positioning Cognitive Workspace as a foundational shift from mere information retrieval to genuine cognitive augmentation in LLMs. This advancement has important implications for the future development of LLMs, suggesting that integrating active memory management and cognitive principles can substantially enhance model performance and efficiency. It recommends further exploration of metacognitive mechanisms and hierarchical memory structures to continue improving context management in AI systems.
👉 Pročitaj original: arXiv AI Papers
