Memory Half Life

An exploratory exponential fit of revisit fidelity against time gap, not a validated measurement of how long a model remembers. It reuses per-pair revisit fidelity:

$$f(\Delta t) = a\, e^{-\Delta t / \tau} + b \tag{1}$$ $$t_{1/2} = \tau \ln 2 \tag{2}$$ $$\mathrm{raw} = \mathrm{clamp}\!\left( t_{1/2} / t_{\mathrm{ref}}, \,0,\,1 \right) \tag{3}$$

The default reference window is $t_{\mathrm{ref}} = 60$ seconds. Fitting uses centered time for numerical stability. Default support checks require at least 5 pairs, more than one time gap, a fidelity range of at least 0.02 and a finite R² of at least 0.3. All observations and the fitted decay time must be finite; the decay time must be positive. Gaps come from native video frame indices / FPS. Insufficient support, no measurable decay or unreliable fits yield not_applicable; NaN/Infinity are never valid scores.

A reliable numeric fit is an official leaderboard result; insufficient support or an unreliable fit remains unavailable. Reports include gap range/span and pair support, but adjacent pairs are not established independent revisit events (event_independence_verified=false). A larger fitted value is not proof of a longer causal memory lifetime. Backend: exponential_fit.

60-Second Same-Input Examples

Audit manifest: docs/assets/metric_examples/longseq_memory_60s_examples_20260909.json.

Input image

60s memory_loop · same benchmark case · gap 0.9904

memory-double_yaw_loop-memory_pool_008

Suite: memory_loopScene: photorealistic / indoor / workspaces

Benchmark prompt target

Turn away and back to the start twice in a row.

prompt source: memory_loop_prompt

The better example still resembles the start after a long loop; the weaker one forgets the place as the time gap grows.

Model result videos will be available with the ModelScope release. / 模型推理视频将在 ModelScope 方案上线后提供。

Higher metric scoreScore 1.0000 · official

Model result videos will be available with the ModelScope release. / 模型推理视频将在 ModelScope 方案上线后提供。

Lower metric scoreScore 0.0096 · official

Input image

60s memory_loop · same benchmark case · gap 0.9939

memory-yaw_loop_180-memory_pool_220

Suite: memory_loopScene: photorealistic / outdoor / terrestrial_landscape

Benchmark prompt target

Turn the camera to the right, then turn back to the starting view.

prompt source: memory_loop_prompt

The better example still resembles the start after a long loop; the weaker one forgets the place as the time gap grows.

Model result videos will be available with the ModelScope release. / 模型推理视频将在 ModelScope 方案上线后提供。

Higher metric scoreScore 1.0000 · official

Model result videos will be available with the ModelScope release. / 模型推理视频将在 ModelScope 方案上线后提供。

Lower metric scoreScore 0.0061 · official

Input image

60s memory_loop · same benchmark case · gap 0.9978

memory-palindrome_loop-memory_pool_406

Suite: memory_loopScene: photorealistic / outdoor / terrestrial_landscape

Benchmark prompt target

Move forward, turn and continue, then retrace the whole path back.

prompt source: memory_loop_prompt

The better example still resembles the start after a long loop; the weaker one forgets the place as the time gap grows.

Model result videos will be available with the ModelScope release. / 模型推理视频将在 ModelScope 方案上线后提供。

Higher metric scoreScore 1.0000 · official

Model result videos will be available with the ModelScope release. / 模型推理视频将在 ModelScope 方案上线后提供。

Lower metric scoreScore 0.0022 · official

Input image

60s memory_loop · same benchmark case · gap 0.9941

memory-dolly_loop-memory_pool_096

Suite: memory_loopScene: photorealistic / outdoor / city

Benchmark prompt target

Move the camera forward through the scene, then back to where it started.

prompt source: memory_loop_prompt

The better example still resembles the start after a long loop; the weaker one forgets the place as the time gap grows.

Model result videos will be available with the ModelScope release. / 模型推理视频将在 ModelScope 方案上线后提供。

Higher metric scoreScore 1.0000 · official

Model result videos will be available with the ModelScope release. / 模型推理视频将在 ModelScope 方案上线后提供。

Lower metric scoreScore 0.0059 · official

Input image

60s memory_loop · same benchmark case · gap 0.9894

memory-double_yaw_loop-memory_pool_181

Suite: memory_loopScene: stylized / outdoor / aquatic_landscape

Benchmark prompt target

Turn away and back to the start twice in a row.

prompt source: memory_loop_prompt

The better example still resembles the start after a long loop; the weaker one forgets the place as the time gap grows.

Model result videos will be available with the ModelScope release. / 模型推理视频将在 ModelScope 方案上线后提供。

Higher metric scoreScore 1.0000 · official

Model result videos will be available with the ModelScope release. / 模型推理视频将在 ModelScope 方案上线后提供。

Lower metric scoreScore 0.0106 · official