Novelty vs Context-as-Memory and WorldMem
Table 1: Design comparison with Context-as-Memory and WorldMem.
| training sample | retrieval | loss covers | |
|---|---|---|---|
| Memorizon | span m ~ U[9, m_max] chunks; history never tokenized | per-chunk top-6 by frustum overlap minus pose distance; union ≤ 60 slots, shared RoPE index | last 10 chunks |
| Context-as-Memory | [VERIFY] | [VERIFY] | [VERIFY] |
| WorldMem | [VERIFY] | [VERIFY] | [VERIFY] |
- Per-chunk retrieval beats one retrieval per block: web-photo Revisit 0.434 vs 0.302.
- A shared RoPE index lets a 10 s-trained model read frames 59 s old.
- "First to decouple" claims in Sec. 1 are reworded to the union-bank mechanism.
Direct comparison
We could not run the released CaM / WorldMem models on our protocol (Appendix C.6). The table above and the per-segment-retrieval row of Table 2 are the closest controlled comparison. [VERIFY]
Evaluation data
[VERIFY and fill: how the 20 web photos were chosen; how the 400 s walks were designed; why existing benchmarks do not cover them.]
Longer horizons
Baselines to 400 s need each model's own pipeline; not run. Fig. 6 shows Memorizon's return intervals beyond 60 s. Stated as a limit.
Abstract notation
Fixed: "bounded by k × K". The sentence before it is split.
Consistency vs quality
Table 13 attributes the quality loss to a train / test mismatch of the bank (clean at training, self-generated at inference), not a weaker generator. [PENDING: G1] reports whether bank noise narrows it.