Archive · 23 Jul 2026
South Korean Researchers Solve AI Forgetfulness with MemEIC
Imagine you have a friend who keeps forgetting old things when they learn new things.
South Korean researchers have found a way to help AI, which has the same problem.
They created a method called MemEIC that lets AI remember both old and new information.
This is important because AI like ChatGPT needs to understand both text and images and keep learning without forgetting.
The researchers tested MemEIC and found it works much better than old methods.
It keeps information separate and only connects it when needed, just like how our brain works.
This could make AI smarter and more reliable.
South Korean researchers developed MemEIC to prevent AI from forgetting old knowledge when learning new information.
MemEIC stores new knowledge in an external memory, retaining old data and improving accuracy.
The technique processes visual and language data separately, connecting them only when necessary.
MemEIC performed 70% better on compositional reasoning tasks compared to traditional methods.
The research was presented at NeurIPS 2025 and involved 1,278 evaluation instances.
- Who
- Researchers from ETRI, POSTECH, and Sungkyunkwan University
- What
- Developed MemEIC, a technique to prevent AI from forgetting old knowledge while acquiring new information
- When
- Presented at NeurIPS 2025
- Where
- South Korea
- Why
- To address the issue of 'catastrophic forgetting' in AI models
Traditional AI Learning Methods
MemEIC Technique
Knowledge Retention
Traditional AI Learning Methods
Traditional methods often overwrite old knowledge when acquiring new information, leading to 'catastrophic forgetting'.
MemEIC Technique
MemEIC retains old knowledge by storing new information in an external memory, preventing loss of previous data.
Handling Multimodal Data
Traditional AI Learning Methods
Traditional methods struggle to integrate visual and language data, often leading to incorrect responses.
MemEIC Technique
MemEIC processes visual and language data separately using adapters and connects them only when necessary, improving accuracy.
- Research Institutions
- Electronics and Telecommunications Research Institute (ETRI), Pohang University of Science and Technology (POSTECH), Sungkyunkwan University
- Conference
- NeurIPS 2025
- Benchmark
- Compositional Continual Knowledge Editing Benchmark (CCKEB)
- Evaluation Instances
- 1,278
- Performance Improvement
- 70% better on compositional reasoning tasks
- Key Researchers
- Lim Soo Jong, Seong Jin