Procedural Knowledge at Scale Improves Reasoning (April 2026)

Reasoning Memory is a retrieval-augmented generation (RAG) framework designed to improve reasoning in language models. It decomposes step-by-step reasoning trajectories into 32 million self-contained subquestion-subroutine pairs, creating a large-scale datastore of procedural knowledge. At inference time, a lightweight in-thought prompt enables the model to retrieve relevant subroutines and apply them as implicit priors. Evaluated on math, science, and coding benchmarks, Reasoning Memory achieves up to 19.2% improvement over no retrieval and 7.9% over the strongest compute-matched baseline. Ablation studies confirm that gains come from broad procedural coverage of the source trajectories and the decomposition and retrieval design.