Autoregressive video diffusion models can generate high-quality frames in real time, but are limited to short clips — push them further, and the KV cache silently discards past context, causing identity drift and quality collapse. We introduce MemRoPE, a training-free framework that solves this with two co-designed mechanisms: Memory Tokens compress evicted frames into evolving dual-rate EMA representations, while Online RoPE Indexing stores keys without positional encoding and applies it dynamically at attention time, keeping temporal aggregation mathematically valid. The result is unbounded video generation with a fixed-size cache — we demonstrate continuous one-hour generation that preserves subject identity and visual fidelity throughout.