Both versions are aligned block by block, in reading order: title, the essentials, then paragraph by paragraph. Where the translation merged or split a paragraph, the matching cell stays empty — we never pair two passages by guesswork.
用AI最让人崩溃的，往往是一个再熟悉不过的瞬间：花半天纠正它，把任务做对，换个新会话，它转头又踩进同一个坑。记忆张量刚刚发布并开源了MemT 1.0，一款“记忆原生”的大模型，专门针对这个毛病。技术报告里有一场测试很能说明问题：用户先给出展览的原定时间，随后敲定周日改期，中间还顺口提了一个未被采纳的周一方案。此后双方又聊了五百多轮、约6.6万token，展览再没被提起。模型在完全不看原始文本的“闭卷”状态下，仍准确报出了原定与改期时间，并指出场地、票价和联系人都没变，把周日的定论和周一的备选分得清清楚楚。
Few things frustrate AI users more than a moment they know all too well: you spend half a day correcting an AI until it gets the task right, then you open a new session and it promptly falls into the same pit again. Memory Tensor has just released and open-sourced MemT 1.0, a “memory-native” large model built specifically to fix this problem. One test in the technical report is especially telling. A user first gives the original date of an exhibition, then settles on rescheduling it to Sunday. Along the way, they casually mention a Monday alternative that was never adopted. The two then talk for more than 500 further rounds, covering roughly 66,000 tokens, and the exhibition is never mentioned again. In a “closed-book” state, with no access to the original text, the model still accurately states both the original and rescheduled times. It also notes that the venue, ticket prices and contact person have not changed, and it clearly separates the Sunday decision from the Monday alternative.
这类场景正是传统方案的短板。市面上多数做法是给模型外挂一个记忆库，把聊天记录存起来，需要时再检索。报告指出，面对跨度这么长的多版本信息，外挂检索很容易把闲聊当成定论。MemT 1.0的思路不同：它把模型的脑力拆成两块，一块是承载通用认知的主干网络，另一块是随交互实时演进的动态记忆状态。每轮对话结束时，模型直接在token级别刷新这块记忆，哪些留下、哪些覆盖，由训练中学出的一套门控机制决定，整个大模型无需重新训练。为了在固定体积的记忆里装下更多细节，团队还引入了Mixture-of-Storage架构，让多位“记忆专家”分头管理存储空间的读写。这些能力靠后天练出来：团队先用200亿token做继续预训练，再用140亿token做后训练，由一个“学霸”教师模型给记忆提取打分。
This kind of scenario exposes the weakness of conventional approaches. Most current solutions attach an external memory bank to the model, store chat logs there, and retrieve them when needed. The report argues that with multi-version information spanning this long a period, external retrieval can easily mistake idle chatter for a final decision. MemT 1.0 takes a different approach. It splits the model's capacity into two parts: a backbone network that carries general cognition, and a dynamic memory state that evolves in real time with each interaction. At the end of every dialogue turn, the model refreshes this memory directly at the token level. A gating mechanism learned during training decides what is kept and what is overwritten, and the entire large model does not need to be retrained. To fit more detail into a fixed-size memory, the team also introduced a Mixture-of-Storage architecture, in which several “memory experts” take charge of reading and writing the storage space. These abilities are acquired through training. The team first ran continued pretraining on 20 billion tokens, then post-training on 14 billion tokens, with a “top-student” teacher model scoring the memory extractions.
在11项系统记忆基准测试中，全程闭卷作答的MemT-1.0-Flash和MemT-1.0-Nano均分分别为50.6和47.0，分别领先同组最强参数化记忆基线12.5分和8.9分。报告还有两个更具体的例子：约1.6万token的历史里，用户问起自己的生日，模型老实回答不知道，没有瞎编；在3.3万token的对话中，模型依据很早之前定下的规则，核对徽章和登记状态，判定两人中一人合规、一人未登记。这些数字全部来自开发方的技术报告，目前还没有独立机构的复现结果。
Across 11 systematic memory benchmarks, the fully closed-book MemT-1.0-Flash and MemT-1.0-Nano average 50.6 and 47.0, respectively. That puts them 12.5 points and 8.9 points ahead of the strongest parameterized-memory baselines in their groups. The report includes two more specific examples. In a history of about 16,000 tokens, when a user asks about their own birthday, the model honestly says it doesn't know rather than making something up. In a conversation of 33,000 tokens, the model applies a rule set early in the exchange to check badges and registration status, and concludes that one of two people complies while the other is unregistered. All these figures come from the developer's technical report, and there are as yet no results from independent institutions to replicate them.
更进一步的测试在真实编程场景中进行。团队先从数万条成功轨迹中提炼出上千条经验，再让模型在遇到报错、空返回或重复无效操作的“卡壳点”时注入经验重跑，据此筛出约4.2万条样本训练记忆模块。在SWE-bench Verified的500项任务上，提前写入经验后，Nano的平均成功率从66.07%升至69.33%，与把同样经验直接塞进上下文的对照组（68.87%）相当。在实战模式下，模型只在报错或原地打转时才当场写入经验，平均成功率达到69.73%，略高于文本经验组的69.13%。覆盖九种编程语言的SWE-bench Multilingual测试中，Nano从60.00%提升到63.56%。这些对比同样出自开发方报告。
The most demanding tests were conducted in real coding scenarios. The team first distilled more than a thousand lessons from tens of thousands of successful trajectories. Then, at “stuck points” where the model encountered errors, empty returns or repeated ineffective actions, it injected those lessons and reran the task. This process yielded about 42,000 samples used to train the memory module. On 500 tasks from SWE-bench Verified, writing experience in advance raised Nano's average success rate from 66.07% to 69.33%, roughly on par with a control group that placed the same experience directly into the context (68.87%). In a live-work mode, the model writes experience only when it hits an error or is going in circles. Its average success rate reached 69.73%, slightly above the 69.13% of the text-experience group. On SWE-bench Multilingual, which covers nine programming languages, Nano rose from 60.00% to 63.56%. These comparisons also come from the developer's report.
对企业来说，最直接的变化是团队经验不必再反复“重教”。报告描述的场景是：项目规范和排错套路，工程师每开一个新会话都要重新交代一遍，企业等于在为重复开销持续买单。在MemT 1.0的设想里，真实业务反馈持续刷新经验库，模型把团队规范逐渐内化，老手少操心，AI少踩老坑。成本账同样关键：在双GPU、单并发的测试条件下，处理64K token历史时，Nano相对完整上下文方案的端到端推理速度约为1.20倍，历史增加到256K时达到约3.99倍；128K历史下，单条交互的持久状态存储约为完整KV cache的六十六分之一。报告还列举了数字分身、慢病管理顾问和个性化私教等长期陪伴场景，但这些仍是设想，文中没有任何落地时间表。
For businesses, the most immediate change is that team know-how no longer has to be taught over and over. The report describes a common situation: project conventions and debugging routines must be explained again every time an engineer opens a new session, so the company effectively keeps paying for this repeated overhead. In MemT 1.0's vision, real business feedback continuously refreshes the experience library, and the model gradually internalizes the team's conventions, so veterans worry less and the AI stumbles into fewer old pitfalls. Cost matters just as much. In tests with two GPUs and single concurrency, when processing a 64K-token history, Nano's end-to-end inference speed is about 1.20 times that of a full-context approach. At a 256K history, the gain reaches about 3.99 times. With a 128K history, the persistent state stored for a single interaction is about one sixty-sixth of the full KV cache. The report also lists long-term companion scenarios such as digital twins, chronic disease management advisers and personalized personal trainers. These remain conceptual, however, and the report gives no timeline for deployment.