Reuse, Don't Recompute: Efficient Large Reasoning Model Inference via Memory Orchestration

TL;DR

ENGRAM-R introduces a memory layer for large reasoning models that significantly reduces token usage and latency by reusing structured memory, maintaining high accuracy in reasoning tasks.

Contribution

The paper presents ENGRAM-R, a novel memory-based inference method that improves efficiency and accuracy in large reasoning models by integrating typed retrieval and compact fact representations.

Findings

Reduces input tokens by 85% and reasoning tokens by 75% on LoCoMo.

Achieves similar efficiency with accuracy gains on LongMemEval.

Demonstrates memory's role in efficient, long-horizon reasoning.

Abstract

Large reasoning models (LRMs) achieve strong accuracy through test-time scaling, generating longer chains of thought or sampling multiple solutions, but at steep costs in tokens and latency. We argue that memory is a core ingredient for efficient reasoning: when evidence already exists, models should think less by reusing structured memory instead of recomputing derivations. We present ENGRAM-R, an inference-time memory layer that integrates typed retrieval with compact fact card representations and explicit citation control. On the LoCoMo benchmark, ENGRAM-R reduces input tokens by 85% and reasoning tokens by 75% compared to full context while maintaining high accuracy. On a multi-hop slice of the LongMemEval benchmark, it achieves similar efficiency with substantial accuracy gains. These results show that memory is not only critical for long-horizon correctness but also a practical lever for efficient reasoning under tight compute, memory, and latency budgets.