From Static Scores to Cognitive Retrieval: A Comparative Evaluation of Memory Prioritization Algorithms for LLM Multi-Agent Systems

Paper 2-1 in the Buddys Architecture Series. A supplementary paper to Paper 2 (DRM, DOI: 10.5281/zenodo.19211620). After three weeks of production deployment of DRM v1 with five LLM agents and over 200 memory entries, we observed score inflation: 54% of entries clustered at scores 9-10, degrading top-5 retrieval accuracy to 34.2%. We evaluated seven retrieval algorithms spanning information retrieval, cognitive science (ACT-R), and control theory (DNFO). The selected architecture — a three-layer cache combining category-level softmax gating, pointer-level multi-dimensional scoring with ACT-R activation dynamics, and on-demand content retrieval — achieved 100% accuracy across all six evaluation scenarios at 2.6 ms per ranking for N=200, using only keyword matching over Markdown files. No embeddings, no vector database, no GPU.