Zhiyu Li, Zeyu Zhang, Linyun Liu, Yiwei Bu, Chenyang Xi, Xichong Zhang et al.
arXiv 2610.26825 66 pages Read the paper
A student that cannot reread its full history risks losing useful context, so training first builds segmented language modeling and then matches a full-context teacher while history is available only through evolving memory.
A single memory block limits online capacity and trainable memory parameters; Mixture-of-Storage expands it into independent experts with shared addresses and separate dense routers for writes and reads.
MemT-1.0 reads accumulated memory during generation, then updates it by scanning the completed interaction token by token with a stability-oriented, symmetrized Gated DeltaNet-2 rule.