Damage to the hippocampal formation permanently disrupts the brain's ability to consolidate short-term episodic experiences into stable long-term memories. Traditional neuroprosthetics rely on fixed, linear attractor matrices that scale poorly in multi-dimensional environments and degrade under upstream signal corruption. This paper presents The Memory Machine, a standalone neuromorphic artificial hippocampus that replaces traditional lookup grids with a scaled dot-product multi-head attention engine driven by an explicit vector path-integration loop. We demonstrate how pairing continuous two-dimensional velocity integration with an attention sequence predictor forces the natural, emergent self-organization of biologically realistic entorhinal grid fields and localized hippocampal place cell responses. Furthermore, we demonstrate how sparse high-dimensional projection allows the network to maintain an 83.6\% retrieval success rate under severe sensory noise injection. This work establishes a complete architectural and mathematical foundation for real-time artificial hippocampal implants.