Agent memory capabilities are becoming one of the key directions in AI research for 2026. However, unlike MemLens, which focuses on the memory of LVLM models themselves, MemEye turns its lens to a different question: How does an Agent's memory system actually perform when operating in a multimodal environment?
MemEye was collaboratively developed by 17 researchers, including Boxuan Zhang, Yihao Quan, and Zeru Shi, and has received 52 upvotes on Hugging Face Daily Papers.
Visual-Centric: Why Vision?
The core design philosophy of MemEye is "visual-centric." The logic behind this is straightforward: in the working scenarios of multimodal Agents, visual information is often the richest yet the most easily forgotten.
Agents need to remember:
- What previously viewed interface screenshots looked like
- Key data from charts shared by users
- The positions of visual elements involved in operational steps
- Image content referenced across multi-turn conversations
Traditional evaluation frameworks are mostly text-centric, overlooking the unique challenges Agents face in visual memory. MemEye fills this gap.
Evaluation Dimensions
MemEye evaluates the memory capabilities of multimodal Agents from multiple perspectives:
- Visual Information Extraction Memory: Can the Agent retain key information extracted from images?
- Visual-Text Associative Memory: Can the association between an image and its corresponding text description be maintained over the long term?
- Temporal Visual Memory: The ability to remember visual information across time sequences.
- Visual Interference Robustness: Whether memory becomes confused when faced with similar but non-identical visual inputs.
Differences from MemLens
On the same day, NVIDIA's MemLens also appeared on Hugging Face. Both focus on multimodal memory, but with different emphases:
- MemLens evaluates the long-term memory capabilities of the LVLM model itself—whether the model can "remember."
- MemEye evaluates the memory module within an Agent system—whether the Agent can effectively utilize memory during task execution.
One is a model-level benchmark, while the other is a system-level framework. The two are complementary.
Why It Matters
As multimodal Agents are deployed in scenarios like customer service, education, and healthcare, evaluating memory capabilities is becoming increasingly important. An Agent that cannot remember what users previously said or images they showed will suffer a significantly degraded user experience.
MemEye provides a set of actionable evaluation tools, enabling developers to quantify an Agent's memory performance and optimize memory modules accordingly.
Developed through the collaboration of 17 researchers and garnering 52 upvotes in the Hugging Face community, this paper demonstrates that this direction is attracting growing attention from researchers.