Proceedings of the Annual Meeting of the Cognitive Science Society

While Large Language Models (LLMs) exhibit impressive fluency, their reasoning often relies on surface-level statistical correlations rather than robust causal understanding. Current reasoning approaches, such as Retrieval-Augmented Generation (RAG), mitigate knowledge obsolescence but typically depend on vector-based retrieval, which mimics associative memory but fails to support the rigorous logical deduction required for complex queries. To address this, we propose a novel framework titled Bridging Associative Memory and Logical Reasoning, which implements a dual-process cognitive architecture. We construct high-fidelity Causal Knowledge Graphs through a human-in-the-loop method, utilizing expert-supervised LoRA fine-tuning to extract reliable causal dependencies from raw text. During inference, a Query Cognitive Planner orchestrates a Dual-Path Memory Access strategy, synergizing associative vector evidence with causal graph traversal. Crucially, a metacognitive validation acts as a "System 2" critic for iterative self-correction. Experiments demonstrate that this cooperative interaction outperforms associative baselines, confirming that explicit "System 2" monitoring effectively reduces logical hallucinations.