MemoLens

Empowering Augmented Reality Glasses with Super Memory

Samiul Alam, Shakhrul Iman Siam, Mi Zhang

MemoLens Team

The Ohio State University

Abstract

With augmented reality glasses, spatial computing can turn everyday experience into searchable memory. MemoLens focuses on super memory: accurately recalling the objects and people a wearer paid attention to or interacted with in the physical world.

We make this possible with two key ideas. First, we use eye gaze captured by AR glasses to identify attended visual regions and create compact memory snippets through gaze-aware spatio-temporal token compression. Second, we organize those snippets in a hierarchical memory structure so user prompts can retrieve relevant memories efficiently over long spans of experience.

We implemented MemoLens with Meta Aria AR glasses and evaluated it on more than 100 hours of egocentric video collected in real-world settings, where it supports accurate real-time retrieval under substantial compression.

MemoLens overview showing gaze-aware memory creation and hierarchical retrieval
0 hours of egocentric AR video
0 aggressive token reduction studied
0 median response TTFT
0 2-hour memory at 94% reduction

System

MemoLens treats gaze as a proxy for attention. Incoming egocentric video is sampled, patchified, aligned with eye-gaze signals, and compressed into visual memory snippets that preserve the regions most likely to matter for later recall.

Gaze-Aware Input

We stream egocentric video and eye gaze from Meta Aria glasses to a paired smartphone for local preprocessing.

Compact Snippets

We merge redundant spatio-temporal visual tokens while retaining attended details.

Memory Tree

We organize short-term snippets into a hierarchy where upper levels summarize longer stretches of experience.

Top-Down Search

We first match user prompts to coarse context, then refine toward the clips needed for response generation.

Memory Creation

  1. 1Sample Frames. Short egocentric clips are sampled into frame-level visual inputs.
  2. 2Align Gaze. High-frequency gaze points are projected onto each frame.
  3. 3Score Patches. Gaussian heatmaps turn gaze into patch-level importance scores.
  4. 4Merge Tokens. Less important visual tokens are merged while attended details remain.

We sample frames from short egocentric clips and project high-frequency gaze points onto the image plane. A Gaussian heatmap spreads each gaze point into a dense importance signal, then patch-level scores guide which visual tokens should be preserved or merged.

This design preserves the parts of the scene that the wearer actually attended to, while reducing the storage and compute footprint required for always-on AR memory.

Results

Below, we summarize what we observe in the primary result figures.

Retrieval performance of MemoLens vs no-token merging

Top-3 retrieval accuracy and storage footprint under token reduction.

X-Clip retrieval performance of MemoLens against no-token merging
X-Clip
CLIP4Clip retrieval performance of MemoLens against no-token merging
CLIP4Clip
MemoLens maintains near-baseline accuracy at low to moderate token reduction. X-Clip stays above 96% top-3 accuracy through 46.9% token reduction, and CLIP4Clip stays above 96% through 62.5%. At 93.8% token reduction, both backbones remain above 78% top-3 accuracy while the plotted two-hour storage footprint is under 1 GB.

Overall performance of MemoLens

X-Clip token merging comparison
X-Clip: token merging comparison
X-Clip impact of gaze on token merging
X-Clip: impact of gaze
LLaVa-NeXT-Video generation consistency
LLaVa-NeXT-Video: generation consistency
CLIP4Clip token merging comparison
CLIP4Clip: token merging comparison
CLIP4Clip impact of gaze on token merging
CLIP4Clip: impact of gaze
Llama-3.2-11B-Vision-Instruct generation consistency
Llama-3.2-11B-Vision-Instruct: generation consistency
MemoLens outperforms intra-frame and cross-frame token merging baselines at all token reduction ratios. At 93.75% reduction, X-Clip reaches 84.07% accuracy, which is 29.17% higher than intra-frame merging and 26.27% higher than cross-frame merging. For the gaze ablation at 94% reduction, X-Clip reaches 78.4% with gaze versus 64.1% without gaze, and CLIP4Clip reaches 78.2% with gaze versus 60.9% without gaze.

Citation

@inproceedings{alam2026memolens,
  author    = {Alam, Samiul and Siam, Shakhrul Iman and Zhang, Mi},
  title     = {{MemoLens}: Empowering Augmented Reality Glasses with Super Memory},
  booktitle = {Proceedings of the ACM International Conference on Mobile Systems, Applications, and Services},
  series    = {MobiSys '26},
  year      = {2026},
  note      = {To appear}
}