Researchers build MEMO, a framework that expands LLM memory without retraining

Researchers from the National University of Singapore (NUS), MIT and A*STAR have presented MEMO, a modular framework that lets language models learn new knowledge quickly without retraining the base model.

The problem of scaling knowledge

Modern large language models are trained on huge datasets and store what they learn in their parameters. When new knowledge has to be added, such as data from a corporate database, medical reference works or up-to-date information on recent events, the traditional approach requires retraining the model from scratch. That means months of computation, millions of dollars in costs and enormous energy consumption. Retraining can also cause the base model to "unlearn" what it already knew, a phenomenon known as catastrophic forgetting.

How MEMO works

MEMO proposes an elegant solution: a separate trainable module called the "memory model." Instead of touching the parameters of the base LLM, new knowledge is encoded into this dedicated module, which works as an additional memory layer and extends the model's capabilities without rebuilding it.

MEMO's architecture consists of several modular components:

• Base LLM (frozen): generates text as usual, with no changes to its parameters
• Memory model: a compact trainable module that stores new facts and knowledge from the corpus
• Integration module: connects the memory to the base model when generating responses for the user
• Retrieval mechanism: finds relevant facts in memory at inference time to provide context
• Training pipeline: updates only the memory module's parameters, leaving the LLM untouched

Practical benefits

This approach gives companies and researchers several key advantages. First, it sharply cuts compute costs: only the compact memory module needs retraining, not a multibillion-parameter LLM with hundreds of billions of parameters. Second, the base model stays stable: its original behavior and previously acquired knowledge are not diluted when new facts are added. Third, knowledge can be updated in hours or days rather than months of painstaking retraining cycles.

For enterprise use, this means being able to adapt off-the-shelf LLMs to specific needs cheaply and quickly: adding domain-specific knowledge, updating information in real time in response to new events, and creating personalized variants of models for different products and customers.

What this means

MEMO is another important step toward modular architectures for AI. Rather than treating a large language model as an indivisible black box that must be fully retrained with every knowledge update, the researchers show that memory and core capabilities can be separated. This opens the way to more flexible, cheaper and more efficient methods of developing and adapting language models.