Sandia Develops AI That Learns Without Forgetting

Sandia National Laboratories researchers are testing a system designed to keep prior knowledge intact while learning new information.

Updated on Nov. 10, 2026 in Artificial Intelligence

Photograph of weathered sensor units on a steel rack in an open New Mexico desert yard under warm late-afternoon light, with a distant mesa. — Sandia National Laboratories is testing an AI system designed to retain prior knowledge while learning new information, in a three-year project led by Tian Ma. AI Illustration. Upload story photo >

Live Poll

Would you trust AI systems that keep learning on their own after deployment?

Sandia National Laboratories is conducting a three-year research project called Adaptive Intelligence for Global Monitoring Lite (AIGM Lite), overseen by Tian Ma. The approach targets catastrophic forgetting, in which AI systems lose accuracy on earlier tasks after learning new information.

Why it matters

Most machine learning systems are updated by collecting new data, retraining and redeploying, a process that can lag behind real-world changes and reduce reliability in fast-changing environments. For New Mexico, home to Sandia, the work points to research capacity in adaptive AI for analytical systems.

In controlled trials simulating gradual data arrival, AIGM Lite reached approximately 90% accuracy compared with roughly 70% for traditional incremental-learning methods. Results from real-world deployment remain unreported.

The players

Sandia National Laboratories

A federally funded research and development center in Albuquerque, New Mexico, known for national security and advanced engineering research.

Tian Ma

The researcher overseeing the AIGM Lite project at Sandia National Laboratories.

The details

The AIGM Lite system uses a dual-system model with a rapid analysis component and a slower component that evaluates new information and prepares updates, borrowing its concept from Daniel Kahneman's 'Thinking, Fast and Slow.' It flags information outside its existing knowledge base for further review rather than making automatic assumptions, and it is being developed mainly for sensor-based analytical systems rather than conversational AI.

The AIGM Lite research project is scheduled to run for three years.

The Tech Race

The project addresses catastrophic forgetting, a problem documented in neural network research since the late 1980s that has resisted incremental-learning fixes. AIGM Lite's dual-system design offers a different route than standard retrain-and-redeploy pipelines used across the industry.

The research aims at AI that stays accurate as conditions change, which could matter for monitoring and analysis tools. For everyday users, no product changes or access details have been announced.

The takeaway

Adaptive AI that retains earlier knowledge could make analytical systems more dependable as data shifts over time. Watch for published results beyond controlled trials to judge how well the approach performs in practice.