Abstract
Graph Retrieval-Augmented Generation (GraphRAG) has advanced multi-hop reasoning but remains bound by the "Static Indexing Paradigm", assuming knowledge is fixed and relationships are explicit. This rigidity fails in dynamic contexts, particularly Counterfactual Reasoning (e.g., "What if X didn't happen?"), where static retrieval fetches conflicting "ground truth." Furthermore, current methods suffer from a Resolution Mismatch, where coarse-grained chunk retrieval introduces semantic noise. We introduce Transient-RAG, a neuro-symbolic framework inspired by Bartlett's Reconstructive Memory. Unlike global graph methods (e.g., HippoRAG), Transient-RAG constructs a query-specific, ephemeral graph at inference time. It employs SemanticSentence Personalized PageRank (SSPPR) to filter noise and injects Hypothetical Edges to bridge logical gaps. Experiments on HotpotQA, 2WikiMultiHopQA, and ifQA show Transient-RAG achieves SOTA performance, significantly improving retrieval accuracy and counterfactual consistency while eliminating index maintenance. Index Terms-Neuro-Symbolic Reasoning, Graph RAG, Counterfactual Reasoning, Inference-Time Compute, Reconstructive Memory I. Introduction In his seminal work Remembering (1932), Sir Frederic Bartlett argued that human memory is not a passive repository of fixed records but an active, imaginative process of reconstruction [1]. This cognitive flexibility enables humans to dynamically assemble knowledge in response to a specific query or premise, filling in gaps through inferences derived from prior experiences. Such an approach contrasts sharply with the architecture of current retrieval-augmented systems, where knowledge is stored statically and fixed prior to inference. Despite the remarkable success of Graph Retrieval-Augmented Generation (GraphRAG) systems, such as HippoRAG [2] and Microsoft GraphRAG [3], they remain inherently limited by the assumption that knowledge graphs are pre-computed and static, thus unable to adapt flexibly to dynamic or counterfactual reasoning tasks. Traditional RAG systems, which rely on the retrieval of large pre-built knowledge graphs, struggle in settings where the query deviates from factual reality.