Record daily life
Smartphones and action cameras capture indoor and outdoor activities. The retained corpus spans 20 participants and 738 clips.
20–120 days per participant
Preprint · 2026Long-term egocentric memory
EgoMonthA Month-Level Egocentric Video Benchmark for Long-Term Spatiotemporal Memory
What can a model remember after a month of everyday life?
Test how it retains, retrieves, and connects experiences across days.
1Nanjing University2Huawei Technologies Co., Ltd.3Tianjin University
* Equal contribution · ✉ Corresponding authors: Lanjun Wang, Zili Yi
301 hEveryday first-person video738 clips · 20 participants
20–120Days per participantLongitudinal recording span
1,443Human-crafted questionsFour-option multiple-choice QA
14Tasks, three cognitive levelsPatterns → episodes → reasoning
Remembering a month of daily life means keeping track of recurring places, sparse events, and subtle changes. A useful memory must distinguish what usually happens from what happened at a particular time, then connect observations when a question requires more than one episode.
EgoMonth brings this challenge to first-person video understanding. Its recordings follow independent daily routines across 20–120 days. The benchmark evaluates both questions grounded in a single recording and questions that combine evidence across multiple videos.
22.4 ppThe reported gap between Gemini 2.5 Pro and the human reference on macro-average accuracy: 71.8% vs. 94.2%.
View paper results · Explore the comparison ↓
View paper results · Explore the comparison ↓
Inside the benchmark
Experience with continuity.
The same participant returns to familiar settings over days and weeks, with important evidence scattered among ordinary routines.
Keep usable context
Screen recordings for quality, viewpoint, and continuity. Anonymize sensitive regions and check the results manually.
18,072 minutes of retained video
Write evidence-based QA
Annotators inspect events, object states, and temporal dependencies to write questions with four plausible answer options.
Fully human-written questions
Verify the evidence
Groups of three reviewers check answer correctness and distractors, revisiting the recordings when they disagree.
1,443 QA pairs across 14 tasks
Recording span describes the interval across days, not uninterrupted 24-hour footage. Average clip length is approximately 24.5 minutes. Dataset construction and statistics.
Explore the recording and activity distributions
Questions in context
Find the evidence.
Connect the experience.
Two original examples from the paper illustrate the difference between following one activity and finding a pattern across recordings.
Procedure planning · Level 3
How did the wearer prepare the ingredients?
The answer depends on the order of several actions. Retrieve the relevant moments in a cooking recording and assemble them into the observed procedure.
Habit inference · Level 1
Which leisure habit recurs across days?
A single clip cannot establish a persistent routine. Combine observations from multiple recordings to identify the behavior that repeats across different settings.
14 tasks / 3 levels
From patterns
to precise recall.
A progressive hierarchy separates stable behavioral patterns, retrieval of specific episodes, and reasoning that depends on multiple pieces of evidence.
Schema Consolidation2 tasks · 154 questions
Repeated experiences provide multiple chances to recover a stable pattern. These tasks ask whether a model can infer regularities across weeks of daily life.
Habit Inference
Identify routines supported by repeated observations across weeks.
138 questions · 9.6% of benchmarkPersonality Inference
Infer enduring behavioral traits from the recorded evidence.
16 questions · 1.1% of benchmarkEpisodic Indexing6 tasks · 873 questions
Find a specific detail, place, object state, or point in time. Similar-looking episodes make precise retrieval harder than remembering the general story.
Detail Retrieval
Recall a fine-grained visual detail from a particular episode.
287 questions · 19.9% of benchmarkSpatial Relation
Recover how objects are positioned relative to one another.
51 questions · 3.5% of benchmarkSelf-localization
Identify the wearer’s location at a specified moment.
85 questions · 5.9% of benchmarkTemporal Ordering
Arrange observed events in their correct chronological order.
127 questions · 8.8% of benchmarkEvent Time
Locate when a particular event happened in the recordings.
117 questions · 8.1% of benchmarkObject Location
Retrieve where an object was last observed.
206 questions · 14.3% of benchmarkCascading Reasoning6 tasks · 416 questions
Combine multiple retrieved observations into a count, route, procedure, or spatial relation. A missed observation or incorrect intermediate state can change the final answer.
Procedure Planning
Connect the steps of a multi-stage activity in sequence.
126 questions · 8.7% of benchmarkEvent Counting
Accumulate occurrences of an event across the evidence.
125 questions · 8.7% of benchmarkObject Counting
Count object instances over the full video context.
57 questions · 4.0% of benchmarkRoute Reasoning
Reconstruct a route from successive places and movements.
42 questions · 2.9% of benchmarkCross-view Spatial Reasoning
Relate spatial evidence across viewpoints and recording days.
30 questions · 2.1% of benchmarkDirection Judgement
Determine orientation from the wearer’s changing viewpoint.
36 questions · 2.5% of benchmarkTask definitions: Table 1. Question counts: Table 5.Question counts · CSV ↓
View the full task distribution
Figure 3(b) · QA distribution across cognitive levelsEnlarge ↗
Reported results
A persistent
memory gap.
Twelve evaluated models, fourteen tasks, and a human reference. Explore overall performance or compare a specific memory skill.
11 open-source models1 closed-source model4 answer optionsDirect answer matching
Paper results · Accuracy (%)
Higher is better
Macro-average accuracy
Dashed line: 25% random chanceHuman: mean of three annotators
Gemini 2.5 Pro: 71.8% · Human reference: 94.2% · Gap: 22.4 percentage points.
Every task, every evaluated model
All values below reproduce the paper’s Table 3, including its reported aggregate scores and model input configurations.
Scroll sideways to compare all models →
| Task / metric | Chat-UniVi-V1.5 | LLaVA-NeXT-Video | MiniCPM-V 4.5 | Qwen2-VL | Qwen2.5-VL-32B | Qwen3-VL-8B | Qwen3-VL-30B-A3B | ShareGPT4Video | ST-LLM | VideoLLaMA3 | VITA-1.5 | Gemini 2.5 Pro | Human |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Input frames | 256 | 64 | 256 | 256 | 256 | 256 | 256 | 64 | 256 | 512 | 16 | 1fps | — |
| Parameters | 7B | 7B | 8B | 7B | 32B | 8B | 30B | 8B | — | 7B | 7B | — | — |
| Level 1 · Schema Consolidation | |||||||||||||
| Habit Inference | 58.7 | 72.5 | 82.6 | 75.4 | 81.9 | 75.4 | 81.9 | 50.0 | 65.2 | 69.6 | 78.3 | 84.8 | 97.8 |
| Personality Inference | 50.0 | 62.5 | 68.8 | 56.2 | 62.5 | 68.8 | 37.5 | 50.0 | 56.2 | 50.0 | 56.2 | 81.3 | 93.8 |
| Level 2 · Episodic Indexing | |||||||||||||
| Detail Retrieval | 34.1 | 38.3 | 66.5 | 58.9 | 61.7 | 54.4 | 56.8 | 34.8 | 34.8 | 55.4 | 51.6 | 77.7 | 95.1 |
| Spatial Relation | 56.9 | 49.0 | 58.8 | 54.9 | 56.9 | 51.0 | 51.0 | 51.0 | 49.0 | 49.0 | 54.9 | 86.3 | 94.1 |
| Self-localization | 29.4 | 41.2 | 54.1 | 52.9 | 49.4 | 52.9 | 52.9 | 31.8 | 35.2 | 43.5 | 48.2 | 82.4 | 97.6 |
| Temporal Ordering | 40.9 | 45.7 | 70.1 | 59.8 | 64.6 | 55.9 | 62.2 | 36.2 | 40.1 | 62.2 | 57.5 | 75.6 | 95.3 |
| Event Time | 44.4 | 46.1 | 57.3 | 42.7 | 47.9 | 41.9 | 50.4 | 45.3 | 42.7 | 41.9 | 49.6 | 60.7 | 95.7 |
| Object Location | 51.9 | 50.5 | 61.6 | 61.6 | 64.1 | 55.3 | 57.3 | 44.7 | 58.2 | 54.9 | 51.5 | 67.0 | 96.1 |
| Level 3 · Cascading Reasoning | |||||||||||||
| Procedure Planning | 46.0 | 29.4 | 68.2 | 68.2 | 78.6 | 71.4 | 70.6 | 30.2 | 46.0 | 60.3 | 66.7 | 81.7 | 94.4 |
| Event Counting | 8.8 | 16.8 | 40.0 | 36.8 | 43.2 | 38.4 | 34.4 | 16.8 | 8.0 | 40.8 | 41.6 | 52.0 | 94.4 |
| Object Counting | 33.3 | 29.8 | 38.6 | 52.6 | 50.9 | 31.6 | 47.4 | 29.8 | 31.6 | 49.1 | 40.4 | 73.7 | 94.7 |
| Route Reasoning | 35.7 | 23.8 | 35.7 | 52.4 | 52.4 | 28.6 | 45.2 | 30.9 | 28.6 | 45.2 | 40.5 | 64.3 | 90.5 |
| Cross-view Spatial Reasoning | 23.3 | 30.0 | 43.3 | 40.0 | 50.0 | 46.7 | 46.7 | 26.7 | 20.0 | 40.0 | 33.3 | 60.0 | 93.3 |
| Direction Judgement | 38.9 | 33.3 | 38.9 | 50.0 | 47.2 | 47.2 | 47.2 | 41.7 | 27.8 | 41.7 | 47.2 | 58.3 | 86.1 |
| Avg · macro | 39.5 | 40.6 | 56.0 | 54.5 | 58.0 | 51.4 | 53.0 | 37.1 | 38.8 | 50.3 | 51.3 | 71.8 | 94.2 |
| Acc · micro | 39.9 | 41.7 | 60.6 | 57.0 | 60.8 | 53.7 | 56.7 | 36.9 | 40.8 | 53.1 | 53.6 | 72.6 | 95.1 |
Source: Table 3, arXiv v1. Models use different input budgets, following their evaluation configurations; this is not a controlled comparison at a shared frame budget. Avg is the reported macro-average across tasks; Acc is accuracy across all questions.
How answers are scored
Outputs are compared directly with the reference answers; no LLM judge is used. Four-option questions have a 25% chance baseline. The paper evaluates each model in a single pass without ensembling.
How the human reference is measured
Three trained annotators, independent of QA creation, answer the full benchmark with unrestricted viewing and replay. Their reported average is 94.2% macro and 95.1% micro, with Fleiss’ κ = 0.78.
What the benchmark reveals
Recognition is only
the beginning.
The reported results point to weaknesses in keeping time, counting events, and joining spatial evidence over long contexts.
Locate the right moment
On Event Time, Gemini 2.5 Pro scores 60.7%, compared with 95.7% for humans. Remembering the event’s general meaning does not guarantee a correct temporal index.
Keep an accurate count
Event Counting reaches 52.0% for Gemini 2.5 Pro versus 94.4% for humans. Multiple open-source models fall below the 25% chance baseline.
Connect different viewpoints
Cross-view Spatial Reasoning reaches 60.0% for Gemini 2.5 Pro versus 93.3% for humans. Stable spatial relationships remain difficult across changing observations.
Gaps above are percentage-point differences calculated from Table 3. The interpretation follows the paper’s analysis.
Scope and interpretation
- The current benchmark includes 20 participants. It does not establish performance across every demographic or environment.
- Coverage extends to 120 days per participant; seasonal and year-long memory are future directions.
- The human reference has unrestricted access to the full recordings, while models use the input configurations reported in the paper.
- Frame-budget and model-size comparisons are observational. They do not isolate the causal effect of adding frames or parameters.
Use EgoMonth
Research access
and resources.
The paper describes a research-only access policy for the benchmark. Metadata and QA pairs can be shared for research; raw videos require additional approval. Contact the authors about availability and access. Redistribution and commercial use are prohibited under the policy described in the paper.
Reference
Cite EgoMonth.
arXiv:2608.13113 · 2026 preprint. The linked PDF includes the full paper and appendices.
@misc{chen2026egomonth,
title = {EgoMonth: A Month-Level Egocentric Video Benchmark for Long-Term Spatiotemporal Memory},
author = {Chen, Weitao and Hu, Jiaxin and Xie, Tianyidan and Li, Yang and Qian, Yuyi and Xu, Banghao and Tang, Ziheng and Wang, Shenyi and Yu, Mingyue and Li, Duo and Shi, Jiacheng and Wang, Gao and Xu, Zhan and Qiu, Zhicheng and Li, Xuanfu and Yang, Jian and Wang, Lanjun and Yi, Zili},
year = {2026},
eprint = {2608.13113},
archivePrefix = {arXiv},
primaryClass = {cs.CV},
doi = {10.48550/arXiv.2608.13113},
url = {https://arxiv.org/abs/2608.13113}
}