Holographic Memory Fails Zero-Shot Knowledge Graph Reasoning

Researchers explored the application of Holographic Reduced Representations (HRR) and phase-only Fourier HRR (FHRR) as a form of holographic memory for knowledge graph embedding (KGE) models. The theoretical promise of HRR lies in its ability to bind and unbind symbols via circular convolution, which could enable zero-shot compositional queries—answering multi-hop questions with relation chains unseen during training. The study evaluated two variants of holographic memory, HRR and FHRR, combined with a modern Hopfield cleanup mechanism, on the FB15k-237 dataset. The findings indicate that while both HRR and FHRR are competitive for single-hop link retrieval, their performance for zero-shot compositional queries remains at chance levels. A detailed mechanistic analysis revealed that the failure is not due to the bind-unbind algebra or the cleanup process itself, as the memory can recover intermediate entities with high fidelity. Instead, the core issue is that facts within compositional chains are inherently harder for the superposed memory to retrieve, even as standalone atomic queries. This suggests a fundamental capacity and interference limitation when multiple facts are stored in superposition, rather than a problem with the compositional mechanism.