Ordo-M: Address-Routed External Memory for Frozen Language Models

Abstract

Ordo-M is a research project exploring external, address-routed memory for frozen large language models. Instead of modifying model weights, knowledge is stored in a sparse external memory and retrieved during inference through learned semantic addressing. The project investigates whether this approach can enable efficient knowledge updates, reduce interference between stored facts, and improve long-context retrieval while preserving the original capabilities of the base model. Ordo-M focuses on reproducible experiments, transparent evaluation, and practical memory architectures for next-generation language models.