δ-mem: Efficient Online Memory for Large Language Models

Researchers introduced δ-mem, a lightweight online memory module for frozen large language models.

It stores past information in a fixed-size matrix and uses it to correct attention during generation.

The method outperformed strong memory baselines and the frozen backbone, with especially large gains on memory-heavy benchmarks.

δ-mem preserves most general capabilities while improving recall, showing that compact external memory can boost LLM efficiency.