δ-mem: 8×8 memory matrix lifts LLM agent benchmarks 31% without fine-tuning

δ-mem is a memory mechanism that augments frozen large language models with a compact online state updated by associative learning. Researchers propose compressing past information into a fixed-size state matrix—as small as 8×8—and using its readout to generate low-rank corrections to the backbone's attention computation during generation. The approach avoids full fine-tuning, backbone replacement, or explicit context-window extension, instead coupling the memory state directly with the existing attention layer.

On memory-intensive benchmarks the gains are substantial: δ-mem reaches 1.31× the frozen backbone's score on MemoryAgentBench and 1.20× on LoCoMo, while averaging 1.10× across all tasks and 1.15× over the strongest non-δ-mem memory baseline. The authors report that general capabilities remain largely intact, suggesting the mechanism adds memory capacity without degrading the model's core performance. The delta-rule learning update—borrowed from neuroscience models of associative memory—lets the state matrix accumulate information online as the model processes new tokens. The preprint appeared May 13, 2026.