CoG-MeM: A Cognitive-Behavior-Inspired and Logic-Aligned Design for Memory Encoding, Retrieval, and Synthesis

Abstract

We propose CoG-MeM, a cognitive-behavior-inspired memory architecture for LLMs that transcends traditional RAG through a logic-aligned pipeline. CoG-MeM features: (1) Logical Compression, employing a high-precision SFT strategy to condense long-form dialogues into structured ``logical chunks'' that ensure the intact preservation of core logical pillars, such as formulas and regulations, while maintaining strict format integrity; (2) End-to-End Retrieval, fine-tuning the model to map complex queries directly to memory entries; (3) Autonomous Triggering, a mechanism to initiate recall via function calling and generate targeted queries; and (4) Logical Arbitration, a context-aware synthesis process that integrates retrieved knowledge with dialogue history, effectively applying external rules whether they reinforce or override pre-trained parametric priors. As a proof-of-concept, this design demonstrates the potential for logical adaptability, establishing a pathway where new knowledge can be assimilated without further weight updates following the initial fine-tuning phase.