To address the challenge of revisit inconsistency in autoregressive generative rendering, researchers have introduced a training-free method designed to maintain 3D world persistence during long-horizon video synthesis. Current conditional video models often fail to maintain visual consistency when a camera returns to a previously visited location after the relevant context has been evicted from the KV cache, despite consistent conditioning inputs like depth maps. To solve this, the authors propose a "loop-closure" mechanism that leverages existing 3D engine correspondences without requiring post-training. The method utilizes temporal correspondence to retrieve pose-matched historical latent chunks, reintroducing them into the KV cache as memory. Furthermore, it employs spatial correspondence—derived from camera pose and depth reprojection—to bias token-level attention toward geometrically corresponding regions within those retrieved chunks. Evaluated on loop-closure trajectories from the TartanAir and TartanGround datasets to simulate complex real-world scenarios, the proposed approach demonstrates superior revisit consistency compared to existing training-free baselines while maintaining overall video quality. This approach effectively closes the loop in autoregressive generation by grounding latent memory in the underlying 3D geometry provided by the rendering engine.