This metric evaluates whether a revisited viewpoint reproduces the frame the model itself generated earlier at the same estimated pose. Pairs are read from the pose track, so the score measures memory rather than whether the action program was executed symmetrically. It does not verify epipolar geometry or illumination statistics—see memory_geometric_closure and memory_rendering_recovery.
The metric shares one analyze_memory_rollout pass with the other memory metrics. A failed memory_return_gate does not block scoring. If strict revisit pairs are absent, nearest earlier poses satisfying the minimum time gap supply fallback pairs (pair_source=nearest_pose_fallback); gate status and pose distances remain in the details. Revisit pairs are bucketed by time gap and at most K pairs are kept (default max_scored_pairs = 24), always retaining the shortest and longest gaps for the forgetting curve. Only those paired frames are decoded; the full 60-second video is never loaded.
For the k-th pair (Iik,Ijk), DINOv2 yields L2-normalized features z(⋅); cosine similarity is mapped into [0,1]. The structural term is per-channel RGB SSIM, averaged and clamped to [0,1]:
The video score is the mean over the selected pairs:
Sfid=K1k=1∑Ksk(3)
A missing DINOv2 checkpoint degrades only this metric. Geometry and rendering can still score.
In WorldAtlas this algorithm is implemented as the official metric memory_revisit_fidelity (pose_aligned_dino backend). Under the default normalization config, raw and normalized outputs lie in [0,1] with higher is better. On memory_loop it is official when scorable pairs and DINOv2 are available, including after a failed gate. Per-pair scores also supply the exploratory decay fit memory_half_life, which is also an official metric when its fit is reliable; if fidelity is unavailable, half-life is skipped with it. The memory_revisit_consistency probe compares protocol-sampled first and last frames. It does not verify a geometric revisit and uses a different measurement protocol.
