Understand Then Memory: A Cognitive Gist-Driven RAG Framework with Global Semantic Diffusion

ABSTRACT

This paper introduces CogitoRAG, a cognitive-inspired retrieval-augmented generation framework designed to mimic human episodic memory and semantic gist formation. The system constructs a multi-dimensional knowledge graph derived from summarized semantic memory structures rather than raw documents. Through query decomposition and entity diffusion, the model retrieves context by associating relevant entities and relationships. The authors also introduce the CogniRank algorithm to rerank passages using both semantic similarity and graph diffusion signals. The framework demonstrates significant improvements in reasoning-heavy tasks across multiple QA benchmarks.