New AI Architecture Improves Reasoning Without Training Through Memory Control

A new research paper introduces a control architecture designed to enhance AI reasoning capabilities without requiring model weight updates. The approach focuses on managing prompt-injected memory, which can boost performance but presents control challenges regarding when retrieved content should be applied. The method addresses applicability control through uncertainty-based routing, confidence-based selective acceptance, and evidence-based governance of memory banks over time. This training-free protocol demonstrated significant improvements on arithmetic benchmarks, achieving gains of +7.0 points on SVAMP and +7.67 points on ASDiv compared to baseline models. The architecture also showed positive transfer effects on question-answering and agent benchmarks, with consistent performance observed across different checkpoints for main arithmetic tasks. The research was published on arXiv under identifier 2604.18206v1.