MemTensor Open-Sources Native Memory Model MemT-1.0, Available in 36B and 288B Tiers
The MemTensor team has open-sourced MemT-1.0, a model built on a native memory architecture, offering two size tiers—36B Nano and 288B Flash—and has released a 66-page technical report. The model connects a Native Memory Block in parallel alongside the standard Transformer Block, uses an erase gate, write gate, and forget gate to achieve token-level dynamic state updates, and extends memory capacity through a Mixture-of-Storage mechanism. According to the figures in the article, under a 256K history, VRAM drops from 5.06 GiB to 40 MiB, and first-token latency remains at 104–112 ms between 512 and 256K.
Why read this
AI score 73
The original text provides specific figures: under a 256K history, VRAM drops from 5.06 GiB to 40 MiB, and TTFT stays stable at 104–112 ms. These can be used to assess the cost impact of the native memory architecture on long-horizon Agent workloads.
Topics
View original postEnglish text is being prepared.
View original
卧槽兄弟们，这个真的牛逼，感觉大模型诞生至今最憋屈的七年失忆史可能要翻篇了！！ 甚至我想说个暴论，这大概是今年大模型架构演进里最性感、但也最击中行业要害的一次底层突围。 刚刚开源的 MemT-1.0 第一次在 Transformer 底层焊入了一颗真正的生物级海马体， 让 AI 从聊完即忘的失忆金鱼，真正进化成了能把经验沉淀进血液里的生命体。 靠暴力堆显卡、拿几个 G 的 KV Cache 硬撑长文本的旧石器时代，今天被这个开源原生记忆大模型正式砸穿了。 他们放出的 66 页技术报告和最高 288B 的开源权重直接甩出了一组降维打击的工程账单： 把记忆变成前向计算流的原生状态后，256K 超长历史下的显存开销从 5.06 个 G 暴降到恒定的 40 兆， 直接把 KV 缓存砍掉了整整一百二十倍以上，而且首字延迟更是不管历史拉得多长，始终稳稳压在一百毫秒出头。 最先击穿传统路径的，是它彻底重构了模型的前向计算流。 以前大家为了让 AI 记住上下文，只能硬着头皮堆长文本或者外挂向量库，代价就是显存撑爆且丢失多步推理因果，它在标准网络旁并联了一套原生记忆模块，在输入进入的瞬间，擦除门、写入门和遗忘门就像生物海马体一样自动运转，不再保留所有的死文字，而是把信息实时消化成动态的内部记忆矩阵。 更绝的是它在真实智能体场景下带来的降维打击，团队把历史纠错中沉淀的一千多条代码经验，直接写入内部的原生记忆状态，完全不占用任何系统提示词的 Token 额度， 结果在标准代码测试集上，连续三次均通过的硬核指标直接提升了百分之八，稳定性甚至反超了直接塞明文提示词的做法。 这意味着经验不再是一段需要模型每次重新阅读理解的生硬规则，而是直接内化成了它的本能直觉，真是牛逼! 我觉得这才是通往真正的递归自我改进最底层的物理逻辑啊, 参数只定义了模型的出厂下限，但唯有持久内化的原生记忆，才能为智能打开复利增长的空间。 以前我们买的模型是聊完即焚的一次性计算工具，从今天起模型正在变成越用越懂业务、越用越值钱的自进化资产。 大家在做智能体或者长对话应用时，最头疼的是长上下文带来的昂贵显存和卡顿延迟，还是模型转头就忘业务规范的失忆症？ 对我来说，光是把 256K 显存压缩一百二十倍且延迟恒定这一项，就已经能让无数原本算不过账的长程业务直接原地复活了hh~
Body · AI translation
Let me make a bold claim: MemT-1.0, just open-sourced, is very likely the watershed moment when large models move from stateless calculators to self-evolving living entities. So the thing that has always blocked the path to AGI was never insufficient compute—the core reason is that all the hundred-billion-scale large models of the past essentially suffered from severe anterograde amnesia, until today, when MemT-1.0 fully open-sourced the first production-grade native memory architecture.
It did not continue to grind away in the dead end of long context and external RAG; instead, it pioneered the world's first MoS Mixture-of-Storage-Experts architecture,
making memory a native state of the model's forward propagation for the first time, cutting VRAM by a full 120x under 256K history, and letting AI completely bid farewell to the fate of having its brain wiped after every single chat.
In this 66-page technical report, the MemTensor team used two models, 36B and 288B, to prove one thing: when AI learns to dynamically erase and write its internal memory matrix in every interaction like the human brain, 1,040 pieces of hard-won historical experience do not need to be stuffed into the prompt—they can be directly internalized into the model's muscle intuition.
MemT-1.0 comes in two sizes, 36B Nano and 288B Flash. Through Gated DeltaNet-2 and the Mixture-of-Storage mechanism, it internalizes massive interactions into the model's native computational state.
Core architecture and 3 breakthrough points: 1. Model architecture (GDN-2): a Native Memory Block is connected in parallel beside the standard Transformer Block, and through erase gates, write gates, and forget gates (hippocampus-like mechanisms), it achieves token-level dynamic state updates, turning memory into a native state of the forward stream.
2. Capacity expansion (MoS): the first Mixture-of-Storage mixture of storage experts, using dense routing and a unified semantic addressing space to linearly expand memory capacity at extremely low cost.
3. Ultimate metrics (VRAM and latency): under 256K history, VRAM plunges from 5.06 GiB to 40 MiB (a reduction of over 120x); time to first token (TTFT) stays steadily flat at 104–112 ms from 512 to 256K, completely solving long-conversation stuttering.
Applicable scenarios: long-horizon Agent development, coding assistants, multi-turn interaction systems, and enterprise-specific knowledge bases that need continuous compounding growth.