VentureBeat reported on a research proposal called delta-mem from Mind Lab and universities. VentureBeat said delta-mem adds a memory module equal to 0.12% of a backbone model’s parameters.
The article described the approach as compressing historical information into a dynamically updated matrix while keeping the underlying model frozen. It said the goal is to reduce reliance on large context windows or RAG for long-running agent tasks.
VentureBeat reported that delta-mem outperformed alternatives on memory-heavy benchmarks while using far fewer added parameters.
Source: VentureBeat