A new open-source memory layer has been developed to enable long context windows for local LLMs without the heavy VRAM requirements of a standard KV-cache. By storing facts in a flat integer store outside the model's context window, the system achieves O(1) lookup complexity and avoids the massive memory overhead associated with large token counts. This approach allows for a private, offline assistant that functions without requiring a GPU cluster.