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

Researchers Daivik Patel and Shrenik Patel have published a paper introducing ENGRAM-R, a novel inference-time memory layer designed to make Large Reasoning Models (LRMs) dramatically more efficient. The core argument is that for efficient reasoning, 'memory is a core ingredient'—when evidence already exists, models should reuse structured memory rather than wastefully recomputing derivations from scratch. This addresses a critical pain point: while LRMs achieve strong accuracy through techniques like long chain-of-thought or sampling multiple solutions, they do so at 'steep costs in tokens and latency.' ENGRAM-R proposes a systematic solution to this scalability problem.

The technical innovation lies in ENGRAM-R's architecture, which integrates typed retrieval with compact 'fact card' representations and provides explicit control over citations. This allows the system to fetch and apply previously computed knowledge efficiently. The results are substantial: on the LoCoMo benchmark, ENGRAM-R slashed input tokens by 85% and reasoning tokens by 75% compared to a full-context approach, all while maintaining high accuracy. It also showed significant accuracy gains on a multi-hop slice of the LongMemEval benchmark. This demonstrates that a dedicated memory layer is not just useful for long-horizon correctness but is a 'practical lever' for deploying efficient reasoning under tight compute, memory, and latency budgets—a key requirement for real-world AI agent applications.