Memorizon: Training World Models Beyond Their Context Window

Tingting Liao; Xuezhi Liang; Hao Li; Guangyi Liu

cs.CVSep 30, 2026

Memorizon: Training World Models Beyond Their Context Window

Authors: Tingting Liao, Xuezhi Liang, Hao Li, Guangyi Liu

Organizations: IFM, MBZUAI · MBZUAI

Abstract

Streaming world models should render a place consistently across repeated visits. Directly supervising such revisits requires training samples that capture both visits, often spanning minutes. Yet dense attention over the full span incurs quadratic costs, making long-span supervision expensive. Memorizon breaks this coupling: long spans are needed for supervision, but not for attention, since the two visits can share a forward pass without including every intervening frame. A training sample covers a span of any length but is scored only on its last $k$ chunks. Instead of tokenizing the history before them, each scored chunk retrieves its own top- $K$ latents by camera co-visibility, and the union of these requests forms a shared bank. The bank is bounded by $kK$ , so the sequence stays bounded however long the span; at the shortest span the recipe is exactly conventional training. Adding the bank raises the cost of a step once; beyond that, a longer span costs little, and going from 100 to 400 s adds 12% to the step time. Against a sliding-window baseline, retrieval raises revisit consistency on every split, and a span long enough to reach the first visit of each return adds a further 24% to 30%, at some cost in image quality; beyond that span, more length no longer helps. Filling the bank from another episode lowers revisit correlation by 83%, so the model uses what it retrieves. Project page: https://tingtingliao.github.io/memorizon

Figures & tables

Panel of Figure 1 — Figure 1: Long-horizon training at bounded cost. (a) Ordinary training covers only a short tail of the episode, so the first visit falls outside it; Memorizon samples a span that holds both visits, scores only its tail and supplies the history through a bank $\mathcal{B}$ of retrieved latents. (b) Bank size against training span: the bank levels off at every $K$ while the candidate pool grows; at $41$ latents there is no history. Shading marks spans beyond $100$ s.

Panel of Figure 2 — Figure 2: The training sequence and what each scored chunk reads. Top: the packed sequence and its RoPE index; bank entries share one index. Grid: the attention mask, one row per scored chunk. Each chunk attends to the first frame, its window of $r{+}1$ chunks and its own top- $K$ : $1{+}K{+}(r{+}1)c=15$ latents, whatever the span. Drawn for $K{=}6$ , $r{=}1$ , $c{=}4$ .

Panel of Figure 3 — Figure 3: Retrieval criterion graded against ground truth. Left: fraction of achievable gain over $960$ scored chunks, from random ( $0\%$ ) to the best candidates ( $100\%$ ); $\lambda=0.1$ and $0.2$ are level within the standard error; bars show $\pm 1$ standard error over chunks. Right: the frame each rule ranks first for one chunk.

	Starting pose				Mid-path				VBench
Method	Revisit $\uparrow$	Gain $\uparrow$	DINO $\uparrow$	D. Gain $\uparrow$	Revisit $\uparrow$	Gain $\uparrow$	DINO $\uparrow$	D. Gain $\uparrow$	Subj. $\uparrow$	Backg. $\uparrow$	Imag. $\uparrow$
LingBot-World 2.0	0.143	$-$ 0.006	0.526	$-$ 0.001	0.147	0.011	0.546	0.017	0.765	0.852	0.741
DreamX-World 1.0	0.146	0.033	0.547	$-$ 0.045	0.109	0.009	0.591	0.007	0.781	0.868	0.684
HY-WorldPlay 1.5 †	0.243	0.040	0.679	0.025	0.194	0.037	0.674	0.027	0.839	0.908	0.702
Matrix-Game 3.0	0.254	0.075	0.717	0.042	0.283	0.089	0.760	0.100	0.813	0.895	0.761
Infinite-World	0.109	0.019	0.610	0.013	0.168	0.070	0.646	0.052	0.789	0.856	0.746

Table 1: Comparison with open world models on the web photographs, with returns to the starting pose and to mid-path poses scored apart. Mean over five seeds; Table 8 gives the spreads. Memorizon : the

100

s run of Table 2 . DINO and D. Gain use DINO ViT-B/16 cosine in place of pixel correlation. † Tracks our walks only to within

1.1

m and

34^{\circ}

Panel of Figure 4 — Figure 4: Comparison with SOTA on a web photograph. Each system is shown late in the walk, when the camera is back near the input view; the input is the reference.

Unseen scenes
	Fidelity to GT		Memory			VBench
Training	PSNR $\uparrow$	LPIPS $\downarrow$	Revisit $\uparrow$	Gain $\uparrow$	DINO $\uparrow$	Subj. $\uparrow$	Backg. $\uparrow$	Imag. $\uparrow$
$10$ s, no retrieval	10.60 $\,\pm$ .06	0.646 $\,\pm$ .005	0.311 $\,\pm$ .018	0.049 $\,\pm$ .020	0.557 $\,\pm$ .020	0.751 $\,\pm$ .006	0.873 $\,\pm$ .001	0.666 $\,\pm$ .014
$10$ s, $10$ -chunk window	10.47 $\,\pm$ .32	0.662 $\,\pm$ .006	0.363 $\,\pm$ .033	0.114 $\,\pm$ .033	0.572 $\,\pm$ .020	0.726 $\,\pm$ .008	0.863 $\,\pm$ .003	0.643 $\,\pm$ .007
$10$ s, top- $K$ , no bank	10.71 $\,\pm$ .20	0.630 $\,\pm$ .005	0.503 $\,\pm$ .022	0.255 $\,\pm$ .026	0.735 $\,\pm$ .011	0.757 $\,\pm$ .003	0.873 $\,\pm$ .003	0.675 $\,\pm$ .016
$100$ s, per-segment retrieval, bank	11.27 $\,\pm$ .22	0.636 $\,\pm$ .013	0.511 $\,\pm$ .040	0.259 $\,\pm$ .023	0.743 $\,\pm$ .032	0.763 $\,\pm$ .003	0.873 $\,\pm$ .002	0.671 $\,\pm$ .022

Table 2: Ablation on training span, bank and retrieval. Rows add one ingredient at a time ( Sec. 4.3 ); spans are drawn uniformly from

10

s to the length given. PSNR and LPIPS need rendered ground truth, which web photographs lack. All runs at step

6{,}000

; mean over five rollout seeds, standard deviation of the per-seed means in small type. DINO is Revisit with the cosine of DINO ViT-B/16 features in place of pixel correlation. Best per split in bold, second underlined. Seen scenes are in Table 7 ( App. C.2 ).

Panel of Figure 5 — Figure 5: Ablation study on five models of Table 2 on a web photograph. Top: back near the input view. Middle: a place a little further along. Bottom: the return to it.

Figure 6: Revisit against return interval over

400

s rollouts on the web photographs; error bars are the standard deviation over five rollout seeds. (a) Revisit per model, grouped by the interval between the two visits;

n

counts returns per seed. (b) Difference from

10

s top-

K

on the same returns. Shading marks the first

60

s, which the tables score.

Appendix figures & tables14 assets

Supplementary material from the paper’s appendix.

Appendix

Translations rescaled by	Pose distance $\mathcal{D}$ ( $w=4$ )	Frustum overlap $\mathcal{O}$
$0.1\times$	$57.1\%$	$39.5\%$
$1\times$ (as posed)	$55.9\%$	$47.4\%$
$10\times$	$46.9\%$	$53.0\%$

Appendix

Table 3: Sensitivity to pose scale. Achievable gain when translations are rescaled with each cue’s constants held fixed.

Per-chunk retrieval $K$	$1$	$2$	$3$	$4$	$6$	$8$	$12$	$16$	$24$
Achievable gain (%) $\uparrow$	$59.6$	$60.0$	$58.9$	$57.4$	$55.4$	$54.2$	$51.1$	$49.1$	$45.9$
Bank capacity	$4$	$6$	$10$	$16$	$24$	$40$	$64$	$\infty$
Achievable gain (%) $\uparrow$	$46.5$	$46.9$	$49.1$	$49.7$	$50.4$	$51.4$	$55.0$	$55.6$

Appendix

Table 4: Retrieval width and bank capacity. Top: each width

K

against its own ceiling, the

K

most similar candidates. Bottom: six frames read from a bank that admits frames by coverage, dropping near-duplicates, and is capped at the given size, together with the latents already completed in the chunk’s window, which supply the rest when the cap is below six;

\infty

lifts the cap but not the de-duplication, so it differs slightly from

K=6

above.

Group	Setting	Value
Model and data	Initialization	Wan2.2-TI2V-5B ( Wan Team, 2025 )
	Camera conditioning	PRoPE ( Li et al., 2025a )
	Video	$864\times 480$ at $16$ fps
Training sample	Latents per chunk $c$	$4$
	Scored chunks $k$ , recent chunks $r$	$10$ , $1$
	Span $m$ (chunks)	$U[10,m_{\max}]$ , $m_{\max}=100$ for Memorizon ( $41$ – $401$ latents)

Appendix

Table 5: Settings. Everything the runs of Table 2 share; the ablation rows differ only as Sec. 4.3 describes.

Definition	Position	Heading	Also required	Used for
Corpus return ( App. B.2 )	$\leq\tau=0.5$	—	camera away $>\varepsilon=1.5$ in between	could one window hold both visits
Evaluation return ( App. C.1 )	$\leq 0.5$	$\leq 15^{\circ}$	$\geq 8$ s apart; away $>1.5$ or $90^{\circ}$	Revisit and Gain
Fold-back ( App. C.1 )	an evaluation return whose path back retraces the path out within $0.5$			camera-following check

Appendix

Table 6: Three definitions of a return. Distances in the units of the corpus (one unit is

4

m).

Figure 7: Revisits in the training corpus. (a) Shortest return interval for three definitions. (b) Share of training windows of each length that contain a return. (c) Per scene, sorted by median, at

\tau=0.5

\varepsilon=1.5

. Dashed lines mark spans of

10

100

200

and

400

Seen scenes, new trajectories
	Fidelity to GT		Memory			VBench
Training	PSNR $\uparrow$	LPIPS $\downarrow$	Revisit $\uparrow$	Gain $\uparrow$	DINO $\uparrow$	Subj. $\uparrow$	Backg. $\uparrow$	Imag. $\uparrow$
$10$ s, no retrieval	11.48 $\,\pm$ .18	0.610 $\,\pm$ .008	0.333 $\,\pm$ .017	0.074 $\,\pm$ .021	0.633 $\,\pm$ .013	0.801 $\,\pm$ .002	0.882 $\,\pm$ .003	0.648 $\,\pm$ .005
$10$ s, $10$ -chunk window	11.08 $\,\pm$ .14	0.641 $\,\pm$ .008	0.321 $\,\pm$ .019	0.081 $\,\pm$ .017	0.599 $\,\pm$ .008	0.774 $\,\pm$ .004	0.875 $\,\pm$ .002	0.635 $\,\pm$ .005
$10$ s, top- $K$ , no bank	11.41 $\,\pm$ .28	0.609 $\,\pm$ .008	0.427 $\,\pm$ .026	0.186 $\,\pm$ .026	0.744 $\,\pm$ .011	0.799 $\,\pm$ .006	0.883 $\,\pm$ .004	0.654 $\,\pm$ .008
$100$ s, per-segment retrieval, bank	11.92 $\,\pm$ .10	0.609 $\,\pm$ .006	0.454 $\,\pm$ .025	0.226 $\,\pm$ .031	0.761 $\,\pm$ .020	0.797 $\,\pm$ .005	0.879 $\,\pm$ .003	0.658 $\,\pm$ .008

Appendix

Table 7: Ablation of Table 2 on seen scenes , held-out episodes of the training scenes along new trajectories. Same runs, protocol and columns; best in bold, second underlined.

	Starting pose				Mid-path				VBench
Method	Revisit $\uparrow$	Gain $\uparrow$	DINO $\uparrow$	D. Gain $\uparrow$	Revisit $\uparrow$	Gain $\uparrow$	DINO $\uparrow$	D. Gain $\uparrow$	Subj. $\uparrow$	Backg. $\uparrow$	Imag. $\uparrow$
LingBot-World 2.0	0.143 $\,\pm$ .026	$-$ 0.006 $\,\pm$ .032	0.526 $\,\pm$ .013	$-$ 0.001 $\,\pm$ .029	0.147 $\,\pm$ .017	0.011 $\,\pm$ .015	0.546 $\,\pm$ .020	0.017 $\,\pm$ .019	0.765 $\,\pm$ .004	0.852 $\,\pm$ .004	0.741 $\,\pm$ .004
DreamX-World 1.0	0.146 $\,\pm$ .009	0.033 $\,\pm$ .021	0.547 $\,\pm$ .016	$-$ 0.045 $\,\pm$ .013	0.109 $\,\pm$ .006	0.009 $\,\pm$ .007	0.591 $\,\pm$ .013	0.007 $\,\pm$ .013	0.781 $\,\pm$ .001	0.868 $\,\pm$ .001	0.684 $\,\pm$ .006
HY-WorldPlay 1.5 †	0.243 $\,\pm$ .014	0.040 $\,\pm$ .034	0.679 $\,\pm$ .022	0.025 $\,\pm$ .031	0.194 $\,\pm$ .011	0.037 $\,\pm$ .010	0.674 $\,\pm$ .004	0.027 $\,\pm$ .004	0.839 $\,\pm$ .002	0.908 $\,\pm$ .001	0.702 $\,\pm$ .008
Matrix-Game 3.0	0.254 $\,\pm$ .018	0.075 $\,\pm$ .017	0.717 $\,\pm$ .013	0.042 $\,\pm$ .019	0.283 $\,\pm$ .012	0.089 $\,\pm$ .020	0.760 $\,\pm$ .016	0.100 $\,\pm$ .016	0.813 $\,\pm$ .004	0.895 $\,\pm$ .002	0.761 $\,\pm$ .003
Infinite-World	0.109 $\,\pm$ .022	0.019 $\,\pm$ .031	0.610 $\,\pm$ .026	0.013 $\,\pm$ .022	0.168 $\,\pm$ .017	0.070 $\,\pm$ .012	0.646 $\,\pm$ .013	0.052 $\,\pm$ .007	0.789 $\,\pm$ .005	0.856 $\,\pm$ .003	0.746 $\,\pm$ .006

Appendix

Table 8: Table 1 with its spreads. Mean over five rollout seeds, with the standard deviation of the per-seed means in small type.

Seen scenes
	Interval between the visits			Type
Training	$8$ – $10$ s	$10$ – $20$ s	$20$ – $60$ s	Start	Mid	All
$10$ s, no retrieval	0.230	0.064	0.073	0.160	0.067	0.078
$10$ s, top- $K$	0.397	0.174	0.185	0.130	0.199	0.191
$100$ s, top- $K$ , bank	0.337	0.216	0.253	0.213	0.252	0.248
Unseen scenes

Appendix

Table 9: Gain by return interval and type. Pixel Gain with the five rollout seeds pooled; seen and unseen scenes as in Table 2 , web photographs as in Table 1 . Start : the earlier frame lies in the first second; mid : every other return. Pooling pairs rather than clips makes all differ slightly from Table 2 . For the

100

s run the three interval bins hold

25

140

and

470

returns on seen scenes,

30

105

and

230

on unseen ones and

20

100

and

605

on web photographs.

Training	Sequence slots	Step time (s)	Wall clock (h)	Peak mem. (GB)
$10$ s, no retrieval	$41$ / $41$	$3.49$	$5.8$	$16.05$
$10$ s, $10$ -chunk window	$41$ / $41$	$4.02$	$6.7$	$16.05$
$10$ s, top- $K$ , no bank	$41$ / $41$	$3.68$	$6.1$	$16.05$
$100$ s, per-segment retrieval, bank	$51$ / $51$	$4.17$	$6.9$	$17.99$
$100$ s, top- $K$ , ordered bank	$65$ / $105$	$8.18$	$13.6$	$28.45$
$100$ s, top- $K$ , bank	$65$ / $105$	$8.02$	$13.4$	$28.45$

Appendix

Table 10: Training cost of the runs in Table 2 , from their logs. Slots: median and maximum over logged steps. Wall clock is the step time over the

6{,}000

steps every run trains for.

System	Correlation $\uparrow$	Direction $\uparrow$	Implied hfov †
Calibration
Rendered frames, engine poses	$0.906$	$0.974$	$89.1^{\circ}$
Open world models
LingBot-World 2.0	$0.855$	$0.966$	$73.3^{\circ}$
DreamX-World 1.0	$0.825$	$0.955$	$132.5^{\circ}$
HY-WorldPlay 1.5	$0.697$	$0.909$	$123.6^{\circ}$

Appendix

Table 11: Camera following : yaw only, web photographs, first

60

s, rollout seed

0

. † Field of view implied by image shift per degree of prescribed yaw; the paths assume

90^{\circ}

. Memorizon : mean over six runs of Table 2 , the three trained on

10

s clips and the

100

200

and

400

s runs.

	Starting pose				Mid-path
Bank contents	Revisit $\uparrow$	Gain $\uparrow$	DINO $\uparrow$	D. Gain $\uparrow$	Revisit $\uparrow$	Gain $\uparrow$	DINO $\uparrow$	D. Gain $\uparrow$
Top- $K$ (ours)	0.377	0.213	0.794	0.218	0.440	0.319	0.826	0.256
Empty	0.176	0.053	0.553	0.085	0.105	0.019	0.504	0.035
$K$ random frames	0.153	0.055	0.583	0.044	0.207	0.109	0.634	0.087
$K$ frames, another episode	0.071	0.012	0.445	$-$ 0.020	0.073	0.016	0.459	$-$ 0.014

Appendix

Table 12: Bank contents at inference. The

100

s run with only the bank’s frames replaced: its top-

K

(ours), nothing,

K

random frames from the same history, or

K

frames from another episode. Web photographs, first

60

s, returns split as in Table 1 , DINO columns as there; rollout seed

0

Panel of Figure 8 — Figure 8: More returns from open world models , extending Figure 4 to three more web photographs. Each block follows one photograph’s walk. Blue tags mark returns to the starting pose; orange tags the two ends of a return to a pose met mid-path. Rollout seed $0$ .

Seen scenes
	Imaging $\uparrow$			Subject $\uparrow$
Training	Generated	GT bank	GT context	Generated	GT bank	GT context
$10$ s, no retrieval	0.648 $\,\pm$ .005	–	0.671 $\,\pm$ .001	0.801 $\,\pm$ .002	–	0.820 $\,\pm$ .000
$10$ s, top- $K$ , no bank in training	0.654 $\,\pm$ .008	0.663 $\,\pm$ .004	0.673 $\,\pm$ .000	0.799 $\,\pm$ .006	0.811 $\,\pm$ .002	0.824 $\,\pm$ .000
$100$ s, top- $K$ , bank	0.646 $\,\pm$ .015	0.666 $\,\pm$ .004	0.668 $\,\pm$ .000	0.804 $\,\pm$ .001	0.818 $\,\pm$ .002	0.820 $\,\pm$ .000
$200$ s, top- $K$ , bank	0.637 $\,\pm$ .012	0.656 $\,\pm$ .003	0.666 $\,\pm$ .000	0.806 $\,\pm$ .006	0.820 $\,\pm$ .001	0.821 $\,\pm$ .000

Appendix

Table 13: Quality against what fills the context on seen and unseen scenes, first

60

s, five seeds. Generated : ordinary rollout. GT bank : the bank holds ground-truth latents. GT context : bank and recent window both ground truth. The ground-truth videos themselves score

0.678

and

0.803

on seen scenes and

0.719

and

0.774

on unseen ones. With a clean history each chunk generates only four latents, so the seed spread rounds to zero.

Panel of Figure 9 — Figure 9: More returns from Memorizon on six web photographs. Blue: a pose close to the first frame’s. Orange and green: two returns, each shown at its first visit and on coming back; they meet the Revisit rule of Sec. 4.1 , or, where a path has only one such return, the same rule at twice its tolerance.

Explore similar work

Sep 30, 2026cs.CV

LOCI: Spatial Linear Memory for Streaming World Models

When a camera revisits a previously observed region, a video world model should reproduce what was there before. This requires both remembering past observations and retrieving the right one for the current viewpoint. Key-value caches preserve visual detail but grow with video length; recurrent memory is compact but compresses history into a fixed-size state, so individual past observations are no longer directly accessible. We introduce LOCI, a hybrid spatial-memory architecture that keeps both representations. In half of the transformer blocks, main attention keeps a key-value cache of past observations; in the other half, it is restricted to the current chunk and complemented by a recurrent linear-attention memory whose reads and writes are conditioned on projective camera geometry, so viewpoint enters both memory addressing and stored content. Recurrent readouts flow into subsequent cache-backed blocks and supply their queries with accumulated scene context. On the public MIND memory benchmark and on held-out recorded trajectories, LOCI reproduces revisited content more faithfully than representative world models and a same-recipe full-softmax model; with full history, it lowers peak memory at equal length by about 30% relative to full softmax. With a bounded bank of retained observations, it streams long videos at constant memory and remains more faithful than full softmax under the same budget.

Ji Xia, Tingting Liao, Xuezhi Liang +2

Institute of Foundation Models, Mohamed bin Zayed University of Artificial Intelligence · Mohamed bin Zayed University of Artificial Intelligence · Pinscreen

Spatial Memory Recurrent Model Video World Models World Models Camera

Ask PDF

Aug 7, 2026cs.CV

Addressable Memory for Video World Models

We study visual persistence in interactive video world models. These models rely on a Key-Value (KV) cache as a growing visual memory to carry forward previously generated frames. However, we find that models can no longer reliably address stored content once rollouts extend beyond the training horizon, because temporal Rotary Positional Embeddings (RoPE) offsets then fall outside the range seen during training and the model struggles to retrieve the relevant visual information through attention. Moreover, naively compressing the cache in the RoPE-rotated space corrupts memory by averaging together incompatible positional phases. To address this, we propose WorldTrace, a training-free memory framework for long-horizon visual persistence. WorldTrace keeps compressed memory addressable by assigning each summary slot a distinct, in-distribution virtual position. Within this addressable cache, we study two memory compression approaches: WorldTrace-Field compresses history for temporal coherence, while WorldTrace-Landmark stores verbatim scene traces at detected transitions for episodic recall. We further introduce LoopBench, a benchmark evaluating whether a compressed cache can reconstruct a previously visited scene after a long detour. WorldTrace-Field improves temporal consistency by +15.5%, and WorldTrace-Landmark improves episodic recall by +19.5% on LoopBench, extending visually persistent generation without retraining.

Xindi Wu, Sven Elflein, James Lucas +5

1NVIDIA · 2Princeton University · University of Toronto +1

Visual Memory Video World Models Peak Memory World Models Rotary Position Embedding

Ask PDF

Sep 21, 2026cs.CV

WorldCrafter: Consistent Video World Model with Implicit 3D-aware Memory

Video world models enable interactive exploration of dynamic environments, yet struggle to respect prior observations over long horizons and across viewpoints. We present WorldCrafter, a video world model that learns a camera-queryable implicit 3D-aware memory for this purpose. The key insight is to let the requested viewpoint shape how multi-view evidence is compressed into the video generator's limited token budget. Trained jointly with the video generator, a memory encoder and pose-conditioned readout module integrate historical observations into a fixed set of target view-specific tokens before denoising, without explicit depth-based correspondences. By combining this memory with recent temporal context and few-step distillation, WorldCrafter enables streaming scene exploration from a single input image or text prompt. Experiments across static and dynamic scenes show substantial gains in long-horizon consistency and camera-control accuracy while preserving visual quality during minute-scale exploration.

Wangbo Yu, Kunhao Liu, Wenbo Hu +8

ARC Lab, Tencent IEG · Peking University

Video World Models Craft Implicit Encoders

Ask PDF