Astrum Verum: Compositional Episodic Cognitive Memory for LLM Agents via Vector Symbolic Architectures

Large Language Models (LLMs) equipped with standard Retrieval-Augmented Generation (RAG) suffer from structural blindness: they cannot distinguish between role-swapped propositions due to the commutative nature of continuous vector search. We introduce a composition-episodic cognitive memory architecture based on Vector Symbolic Architectures (VSA). By decoupling semantic embedding from structural role-binding, the system preserves exact structural relationships and sequential episodic order without sacrificing continuous generalization. We demonstrate that the architecture sustains graceful capacity scaling up to 16,000 facts, achieves near-perfect episodic order recall within a bounded working window, and maintains high semantic fidelity between the original dense embedding space and the high-dimensional space. Our results show that this hybrid neuro-symbolic approach significantly outperforms standard RAG baselines on role-ambiguous queries while providing a mathematically rigorous foundation for verifiable AI memory.