An attention mechanism that shares a single low-rank latent representation across all attention heads instead of maintaining separate keys and values per head.