ETRI Develops MemEIC Technology to Prevent Catastrophic Forgetting in Multimodal AI

The Electronics and Telecommunications Research Institute (ETRI) has unveiled a new AI architecture called MemEIC designed to eliminate "catastrophic forgetting" during continuous learning processes. By utilizing external memory adapters for visual and linguistic data, the technology allows multimodal AI models to update information without degrading existing knowledge or causing hallucinations. This breakthrough is critical for the telecommunications sector, where maintaining up-to-date, accurate data for intelligent services and industrial applications is essential for operational reliability.

Researchers at ETRI, in collaboration with POSTECH and Sungkyunkwan University, presented the Continual and Compositional Knowledge Editing Technology (MemEIC) at the NeurIPS 2025 conference in San Diego. The technology addresses a core limitation in current multimodal models, such as ChatGPT and Gemini, where learning new information often causes the system to forget or mix up previously acquired knowledge, a phenomenon known as "catastrophic forgetting." Unlike traditional methods that modify internal parameters and risk destabilizing the model's existing structure, MemEIC stores new information in external "visual" and "language" adapters. This modular approach ensures that visual and linguistic data are kept independent until a "knowledge connector" links them based on specific context to answer compositional questions accurately.

The effectiveness of the MemEIC system was validated using the Compositional Knowledge Editing Benchmark (CCKEB), which consists of 1,278 items. In sequential editing experiments involving hundreds of pieces of new knowledge, the ETRI-developed technology achieved an accuracy level of approximately 70% in answering complex, compositional questions. This performance is more than double the 36% to 52% range recorded by existing state-of-the-art technologies. Furthermore, the system demonstrated high "locality," a characteristic where the stability of responses to existing questions remains unchanged even after new data is integrated, preventing the degradation of the model's original intelligence.

For the telecommunications and ICT industries, this technology offers a scalable solution for deploying intelligent services that require constant updates, such as real-time policy changes, legal information, and industrial product data. Director Lim Soo Jong of ETRI’s Language Intelligence Research Section noted that the achievement provides a foundation for multimodal AI to reflect up-to-date information required in real-world service environments while maintaining high reliability. By minimizing internal interference and hallucinations, MemEIC allows for more dependable AI-driven customer support and industrial monitoring tools that can adapt to rapidly changing market conditions without losing their core training.

This research was supported by the Ministry of Science and ICT and the Institute of Information & Communications Technology Planning & Evaluation (IITP) as part of the Next-Generation Generative AI Technology Development Program. Lead author Seong Jin explained that MemEIC overcomes the interference problems inherent in conventional methods by storing different types of knowledge independently and connecting them only when needed. As a non-profit government-funded institute, ETRI intends to further advance this technology to ensure it can reliably reflect diverse information from industrial sites, reinforcing Korea's position as a leader in the global ICT and telecommunications landscape.