Brain Agent Supreme (BAS) is a novel cognitive architecture designed to improve information retrieval and reasoning over long documents without relying on large generative language models. Instead of processing an entire document within a single context window, BAS distributes the workload across multiple specialized agents, each responsible for a document segment and connected through a central associative memory system.More specifically, BAS addresses three major challenges in long-context AI systems: information loss in extended documents, temporal fragmentation of facts, and multi-hop reasoning failures. By combining semantic and lexical retrieval, shared memory mechanisms, and deterministic reasoning modules, the system achieved 100% accuracy on the LOCOMO v2 benchmark (520 tests) while operating entirely on CPU and without dependence on generative LLMs.The work also introduces the concepts of Failure Analysis Taxonomy and Calibrated Ignorance, emphasizing safe, evidence-grounded responses over potentially hallucinatory generation. BAS is proposed as a reliable deterministic reasoning substrate for future hybrid AI architectures that combine factual retrieval with lightweight natural-language generation .
Francesco Bulla (Mon,) studied this question.