MemCoT: Test-Time Scaling through Memory-Driven Chain-of-Thought

TL;DR

MemCoT introduces a test-time memory scaling framework for LLMs, transforming long-context reasoning into an iterative search process to improve accuracy and reduce hallucinations.

Contribution

It proposes a novel memory-driven reasoning framework with multi-view long-term and dual short-term memories, enabling more effective long-context reasoning in LLMs.

Findings

Achieves state-of-the-art performance on LoCoMo and LongMemEval-S benchmarks.

Enhances reasoning accuracy by iterative evidence localization and contextual expansion.

Reduces hallucinations and catastrophic forgetting in long-context reasoning.

Abstract

Large Language Models (LLMs) still suffer from severe hallucinations and catastrophic forgetting during causal reasoning over massive, fragmented long contexts. Existing memory mechanisms typically treat retrieval as a static, single-step passive matching process, leading to severe semantic dilution and contextual fragmentation. To overcome these fundamental bottlenecks, we propose MemCoT, a test-time memory scaling framework that redefines the reasoning process by transforming long-context reasoning into an iterative, stateful information search. MemCoT introduces a multi-view long-term memory perception module that enables Zoom-In evidence localization and Zoom-Out contextual expansion, allowing the model to first identify where relevant evidence resides and then reconstruct the surrounding causal structure necessary for reasoning. In addition, MemCoT employs a task-conditioned dual short-term memory system composed of semantic state memory and episodic trajectory memory. This short-term memory records historical search decisions and dynamically guides query decomposition and pruning across iterations. Empirical evaluations demonstrate that MemCoT establishes a state-of-the-art performance. Empowered by MemCoT, several open- and closed-source models achieve SOTA performance on the LoCoMo benchmark and LongMemEval-S benchmark.