Oct 20 — Shanghai-based Memory Tensor team released and open-sourced the
MemT-1.0 series. Unlike models that reset each conversation, MemT-1.0 integrates
persistent, updatable internal memory into model computation so historical
information and task experience can be invoked in later related tasks without
re-supplying original text. The open-source package includes MemT-1.0-Flash
(288B parameters) and MemT-1.0-Nano (36B). The team added native memory modules
to selected Transformer blocks, enabling standard attention and memory attention
to run in parallel and making memory part of forward propagation.
