Researchers proposed Co-LMLM, a limited-memory language model design that externalizes factual knowledge into a knowledge base rather than storing it entirely in model weights. During generation, the system queries that external store as needed.
The approach targets a core problem in LLMs: models can become stale, costly to update, and unreliable when factual knowledge is embedded only in parameters. External memory can make updates and provenance easier to manage.
If practical at scale, limited-memory designs could reshape how developers build retrieval-heavy AI systems and reduce the pressure to pack every fact into ever-larger models.