eMoT: evolving Memory-of-Thought via Symbolic Anchoring and Memory Corrosion

Researchers introduce eMoT (evolving Memory-of-Thought), a framework that enhances LLM reasoning by treating reasoning processes as dynamic, evolving memories rather than static sequences. The system combines memory corrosion mechanisms, symbolic anchoring for deterministic computation, and consistency refinement to reduce hallucinations and improve multi-step reasoning accuracy, achieving 100% on Game of 24 and significant gains on mathematical benchmarks.

eMoT addresses a fundamental architectural limitation in how large language models approach complex reasoning tasks. Rather than treating each reasoning step as an isolated generation event, the framework implements a memory-based system that learns from successful reasoning patterns and discards less effective ones—mimicking human cognitive refinement. This represents a meaningful shift from static prompt engineering toward dynamic, adaptive reasoning structures.

The symbolic anchoring component is particularly significant for practical applications. By delegating numerical computation to deterministic Python execution rather than relying on LLM probability distributions, the framework reduces a primary failure mode in mathematical reasoning. This hybrid neural-symbolic approach has proven effective across multiple benchmarks: Game of 24 (100% accuracy, 17.6% improvement), GSM8K, ASDiv, SVAMP, and MGSM. Notably, these gains were achieved using lightweight backbone models, demonstrating that framework design—not raw model scale—drives the performance improvements.

For the AI development community, eMoT validates an important principle: reasoning reliability scales through architectural innovation rather than parameter scaling alone. This has implications for resource efficiency and deployment constraints, particularly for applications requiring deterministic numerical accuracy. The consistency-driven refinement module's ability to align neural outputs with symbolic computations suggests a path toward more trustworthy AI systems in high-stakes domains like finance and scientific computing.

The framework's applicability to problems requiring symbolic grounding and iterative refinement positions it as a model for future reasoning architectures. Ongoing development will likely focus on scaling the memory corrosion mechanisms and extending symbolic anchoring to broader computational domains beyond arithmetic.

• →eMoT achieves 100% accuracy on Game of 24, surpassing baselines by up to 17.6% through dynamic memory and symbolic grounding.
• →Hybrid neural-symbolic approach delegates deterministic computation to Python, reducing LLM hallucinations in mathematical reasoning.
• →Performance gains demonstrated on lightweight models prove framework design, not model scale, drives reasoning improvements.
• →Memory corrosion mechanism reinforces high-utility reasoning patterns while discarding inefficient ones, enabling adaptive reasoning refinement.
• →Consistency-driven refinement aligns neural inference with symbolic outcomes, reducing logical error accumulation across multi-step tasks.