Semantic-Logic Decoupled Cognitive Architecture —— An Interpretable, Correctable, and Continual Learning General AI System

Current large language models (LLMs) mix semantic knowledge and logical reasoning in the same high-dimensional vector space, leading to unlocatable errors, high correction costs, and mutual contamination between logic and semantics. This paper proposes a new cognitive architecture, whose core is to force the separation of semantic vectors and logical relation vectors, and coordinate them through a "reverse filling" mechanism. This architecture has seven core advantages: ① Physical isolation of semantics and logic; ② Precise error localization and low-cost correction; ③ Interpretable reasoning process; ④ Continual learning without catastrophic forgetting; ⑤ Automatic construction of initial rule bases using open-source LLMs; ⑥ Balancing real-time performance and latency through cached precomputation; ⑦ Supporting one-shot teaching to achieve "mastering new knowledge with one teaching session just like teaching humans". At the same time, it honestly points out its inherent limitations: high latency when the cache is not hit, making it unsuitable for millisecond-level motion control scenarios; noise in the initial rule base; and the need for a small amount of human feedback. This paper explains the necessity of "reverse filling" through two intuitive analogies: Newton's laws and mathematical application problems, elaborates on the engineering implementation strategy of scope division (including rule matching and cache decision-making), compares it with reinforcement learning, and points out that heuristic methods are simpler, more reliable and interpretable. A complete engineering implementation path is given in the paper, and all steps are based on existing open-source technology stacks. To the best of our knowledge, this paper is the first complete cognitive architecture that forcibly physically separates semantics and logic, coordinates them through reverse filling, and is supplemented by cached precomputation and closed-loop iterative correction, providing a new paradigm for building interpretable and evolvable general artificial intelligence.