TASKS INTRODUCED IN THE LTE PAPER
Spatial Memory Benchmark (SMB) tests whether a memory system can retrieve objects by their state history and find their last occurrence across hours of egocentric observations. It is constructed from EgoLife multi-day recordings.
Public benchmark description. Annotations and evaluation code are not released here.
01 / TASK DEFINITIONS
Two complementary tests of object-centric memory, with 300 queries each.
STR / 300 QUERIES
Locate objects matching a description of their motion or state history, within optional semantic spatial regions and temporal constraints.
LOR / 300 QUERIES
Retrieve the last occurrence of a described object inside a specified lookback window, optionally restricted to a spatial region.
Last occurrence means last observed evidence, not a guarantee of an unseen object's current position. These are task definitions, not executable annotation schemas or live inference.
02 / PAPER-REPORTED PROTOCOL
For both tasks, the paper reports success rate: a retrieved frame must lie within the object's ground-truth temporal span and have bounding-box IoU ≥ 0.3. The paper also reports stricter IoU ≥ 0.5 results and per-horizon analysis.
SMB is constructed from EgoLife, whose source collection contains 300 hours from six participants over seven days in a shared home. Individual sessions reach 50 hours. This is the source collection's scale, not the length of every SMB query; LOR looks back at most 24 hours.
The LTE paper also evaluates Ego4D Natural Language Queries (NLQ) and Visual Queries 2D (VQ2D). Those are established Ego4D tasks, not part of SMB. SMB introduces STR and LOR using EgoLife recordings to study long-term embodied agent memory, object state history, and natural-language video retrieval.
Construction and annotation details: §4.1 and Appendix A of arXiv:2609.04802v1. SMB does not evaluate general lifelong learning or robot navigation.
03 / RESULTS IN THE PAPER
Success rate (%), 300 queries per task, with bounding-box IoU ≥ 0.3 inside the ground-truth temporal span.
| Method | STR ↑ | LOR ↑ |
|---|---|---|
| Qwen3-VL-8B + Grounding-DINO | 21.5 | 25.1 |
| Qwen3-VL-235B + Grounding-DINO | 31.9 | 34.4 |
| KFMem (3D-Mem-style) | 19.8 | 33.8 |
| VideoAgent | 24.7 | 30.5 |
| LTE-based system (ours) | 45.3 | 48.7 |
Source: §4 of the paper. These are paper-reported results, not an independently reproduced leaderboard or new measurements from this website.
04 / AVAILABILITY AND CITATION
SMB annotations and evaluation scripts are not released here. No dataset-release date is announced.
The planned ST-Mem memory-core release is separate from benchmark evaluation code. This website does not distribute source video, model weights, or SMB query annotations, and it does not grant dataset access or redistribution rights.
For source recordings, consult the original EgoLife project and its access terms. EgoLife is the work of Jingkang Yang and collaborators (CVPR 2025); Ego4D is the work of Kristen Grauman and collaborators (CVPR 2022).
For SMB, cite Linguistic Trajectory Encoding for Efficient Long-Horizon Spatial Memory in Embodied Agents, by Tianyidan Xie and collaborators, arXiv:2609.04802. There is no separate SMB DOI claimed here.