LLM agent memory at 0.12% of model parameters

Researchers introduced delta-mem, a lightweight memory module for LLM agents that compresses interaction history into a fixed-size matrix.

It keeps the base model frozen, uses only 0.12% of backbone parameters, and outperforms a stronger baseline on memory benchmarks.

The approach could give coding and data-analysis agents cheaper, lower-latency long-term memory without relying so much on huge context windows or RAG.

This may be especially useful in enterprise workflows where retaining and reusing context matters.