Large-Scale Procedural Knowledge Retrieval: A New Paradigm for Enhancing Reasoning Capabilities with Reasoning Memory

Reasoning Memory is a retrieval-augmented generation framework for reasoning models. By retrieving relevant subroutines from 32 million procedural knowledge entries, it allows models to reuse historical reasoning experiences and achieve significant performance improvements in math, science, and coding tasks. Its core lies in addressing the limitations of existing test-time expansion methods, which handle problems in isolation and cannot reuse procedural knowledge, thus bringing a paradigm shift to reasoning models.

Test-time expansion methods (e.g., Chain of Thought, multiple sampling) can improve the accuracy of reasoning models, but they have fundamental limitations: each problem is handled in isolation, making it impossible to systematically reuse experiences from similar historical problems, especially ignoring procedural knowledge (metacognitive skills such as problem decomposition, strategy selection, verification and backtracking), which forces models to build reasoning processes from scratch.

Reasoning Memory focuses on retrieving and reusing procedural knowledge (knowledge of "how to do things", such as problem decomposition, strategy selection, etc.). Its technical implementation includes: 1. Trajectory decomposition: splitting existing reasoning trajectories into 32 million "subproblem-subroutine" pairs to form a knowledge base; 2. Retrieval during reasoning: simulating the reasoning method of human experts reusing strategies through three steps: explicit subproblem formulation → retrieval of relevant subroutines → procedural prior reasoning.

In six benchmark tests including math, science, and coding, Reasoning Memory consistently outperforms comparison methods such as traditional document RAG and full-trajectory RAG: it improves by up to 19.2% compared to methods without retrieval, and by an average of 7.9% compared to the strongest computational matching baseline. Ablation studies show that the key to success lies in: the extensive procedural coverage of source trajectories, and the decomposition and retrieval design of subproblem-subroutine pairs.

Implications: Shift from "memorizing facts" to "learning to solve problems", expand retrieval augmentation to areas such as strategies/methods/verification rules, inspired by human cognition (experts rely on procedural knowledge). Application value: Improve reasoning efficiency (avoid repeated exploration), enhance interpretability (trace decision-making basis), and support continuous improvement (the knowledge base forms a closed loop as it expands with new problems).

Current limitations: High cost of knowledge base construction, retrieval accuracy depends on design, strong domain specificity. Future directions: Dynamic knowledge base update (automatically extract procedural knowledge from new reasoning), cross-domain migration, deep integration of retrieval and reasoning, and building a hierarchical procedural knowledge system.