MRI・PKUTECHが「AI Memory RAG」開発、過去の議論を踏まえた分析・回答が可能に

On July 8, Mitsubishi Research Institute (MRI)—a major think tank and consulting services firm—and PKUTECH, an AI and data analytics company, announced the joint development of "AI Memory RAG," a new retrieval-augmented generation technology designed to overcome limitations in how enterprises extract value from accumulated data.

Traditional RAG technology works by allowing AI systems to supplement their pre-trained knowledge with relevant company-specific data, improving the accuracy of responses. However, this approach has a significant weakness: while it excels at finding similar documents or passages, it struggles to capture the full context of how decisions evolved over time. Documents that continuously accumulate—such as news feeds and meeting minutes—create particular challenges. Past discussions, the reasons why approaches changed, and the processes through which consensus was reached all become difficult to track and retrieve, and temporal changes become hard to discern.

AI Memory has recently emerged as a solution to this problem. Rather than treating each query in isolation, AI Memory systems maintain persistent memory of past information, dialogue histories, and accumulated knowledge, which the generative AI then continuously draws on to produce answers. Recognizing the promise of this approach, MRI and PKUTECH built on AI Memory by adding technology to maintain and search information while preserving both temporal and contextual relationships in appropriate data formats, creating AI Memory RAG.

The key innovation lies in how the system structures and stores data. Documents targeted for search—primarily news and meeting minutes—are decomposed into granular components: pages, chapters, tables, and metadata. The system then automatically determines the optimal storage format for each component based on its content characteristics. Information is held in one of three structural forms: a temporal (time-ordered) structure, a graph or network structure that captures relationships between people and keywords, or a repository structure that tracks change history and revision logs. When answering a query, the system combines AI Memory's query-processing mechanism with these structured representations, analyzing the incoming query to generate a plan for which data to retrieve and what reasoning steps to perform.

MRI and PKUTECH envision the technology being deployed in three main domains: news monitoring and intelligence analysis, development management and meeting minute management, and knowledge transfer across organizations. The companies plan to verify accuracy and conduct proof-of-concept trials to increase practical utility, with the ultimate goal of integrating AI Memory RAG into full commercial solutions.

The development of AI Memory RAG reflects a growing recognition of a practical gap in enterprise AI deployment. While traditional RAG has become a standard approach for grounding language models in proprietary data, it operates primarily as a semantic search tool—retrieving similar documents or passages without full understanding of how information has evolved or changed. Organizations increasingly face a different problem: they need AI systems that can trace back through accumulated institutional knowledge, understand how decisions were made in the past, and apply that historical context to new questions.

AI Memory is an emerging paradigm that addresses this by having AI systems maintain persistent memory of conversations and knowledge over time, mimicking human institutional memory. By combining this with structured data representations—temporal sequences for time-stamped information, graph structures for relationships between people and concepts, and change-history structures for tracking decisions and reversals—MRI and PKUTECH have created a system that can retrieve not just "what was said" but "how we got here." This is particularly valuable in domains like intelligence analysis and development management, where understanding the evolution of thinking and the rationale behind decisions is as important as the decisions themselves.

No comments yet. Be the first to share your thoughts!