KeyMemRT Compiler and Runtime: Unlocking Memory-Scalable FHE

KeyMemRT addresses a critical physical bottleneck in FHE: the massive size of Galois/rotation keys which often exceed available RAM, leading to thrashing or execution failure. While the project currently shows low public engagement (0 stars, 3 forks), it represents high-value academic research (linked to an ArXiv paper from early 2025). The defensibility is currently low due to its status as a research prototype, but the domain expertise required to replicate it is significant. It competes conceptually with specialized FHE compilers like Zama's Concrete or Google's HE-Transformer. The primary moat is the specific scheduling algorithm for key loading, which is a niche but vital optimization for hardware-constrained FHE. Frontier labs are unlikely to build this as a standalone tool, but are very likely to integrate similar logic into their own privacy-preserving stacks (e.g., Google's FHE transpiler) if FHE becomes a production standard for LLM inference.