Structural Associative Retrieval: Zero-Cost Memory Linking via Embedding Clustering and Lexical-Semantic Pipelines

Current memory systems for conversational AI agents rely on embedding similarity for retrieval, which captures topical relevance but fails at associative recall. I propose Structural Associative Retrieval (SAR), a three-stage pipeline that achieves associative memory linking without training new models or calling external APIs. SAR reuses existing embeddings through clustering for coarse structural grouping, applies part-of-speech extraction for skeletal decomposition, and performs word-level embedding comparison for fine-grained ranking within clusters. On a controlled dataset of 50 personal memories, SAR achieves near-perfect associative accuracy. I present ablation experiments across five retrieval variants, demonstrating that each pipeline stage contributes irreplaceably. SAR operates entirely locally with zero API cost and compresses memory indices from 768-dimensional vectors to single-integer cluster IDs (768x reduction)