前言：当记忆不再是被索引的离散数据，而是通过时序训练内化为可推理、可演化的参数化网络时，知识工程的核心命题已从“What is true?”转向“What is ME?”，这是知识图谱从通用认知基础设施迈向个性化心识科学的临界点。本文介绍的内容本质上是一场关于人类认知的实验，而我们对实体关系与动态图谱的探索，或许藏着解锁它的密钥。
Me.bot 使用围绕个人的个性化数据，而不是全人类的泛化数据。在新的社会秩序之中，在我和你见面之前，让代表你精神的 bot 和代表我精神的 bot 先聊一下。
产品初衷 🍅
Mindverse（心识宇宙）创始人陶芳波在一次采访中这样介绍他们的产品 Me.bot（Me.bot - Your AI Second Brain）：
“Me.bot 不仅仅是通过 RAG 或 Long context 对知识进行回溯，而是基于用户记忆为每个人训练个人模型，从而提供启发和个性化、主动式的服务。我们想要离世界远一点，离个人近一点。”
在 ChatGPT 和 Deepseek 争夺通用智能的当下，心识宇宙押的是另一条轴线：不做公用 AI 的性能提升，而是给每个人做一个专属的“me”。它不是照顾起居和信息的助手，而是数字版的人，是人的延伸。整个赌注可以压缩成一句话：比起预测下一个 token，他们想预测你的下一个想法。
这个说法是可以检验的，陶芳波给了一段轶事作证据。心识宇宙一直在训练他个人的 Me.bot；35 岁生日那天他说了一段话，而他的个人模型说出的几乎一模一样：“我是 Mindverse 创始人陶芳波，今天是我 35 岁的生日，今天我回顾了一下，我对 30 岁的那个决定非常感恩，是一个特别正确的决定。”
问到如何做到这一点，他的回答只有一个词：“记忆”。记忆是一个人最重要的数据集，足够密、足够长、并且只属于这一个人；有了它，才谈得上预测一个人想法的河流往哪里流。
Me.bot 与用户共同的记忆 🍅
架构上的分野在这里。传统 AI 的记忆管理（RAG、长上下文）本质是外挂式的数据调用：模型本身保持通用，个性化只存在于被取回、塞进 prompt 的那部分内容里。Me.bot 反过来，把用户记忆参数化，把个人经历和思维模式内化成模型自己的属性。这个设计暗合神经科学“记忆塑造人格”的判断：当 AI 能基于一个人的生命时序数据（听到的、看到的、思考的）预测“下一个想法”时，它就不再是被动响应指令的工具，而在变成用户思维的数字孪生。
关键差异在时序训练框架。大模型的参数一冻结就是几个月，Me.bot 想模仿的是人的成长方式：人睡觉的八个小时会把白天的记忆压进大脑，Me.bot 就每天夜里把白天学到的东西压进参数，模型每日更新。对齐层面的后果，陶芳波说得很直接：“RLHF（Reinforcement Learning from Human Feedback）的意思是向人对齐，让 AI 的利益与人类更接近。而我的个人模型一定要对齐我的个人利益，而不是 generally 全人类的。这是一个从 RLHF 到 RL’ME’F 的转变。”
对主流 Agent 方案（Coze、Workflow 这一类），陶芳波同样不客气：它们本质是“脚手架式外挂”，靠外部逻辑把智能拼装起来。现有 Agent 用拆解任务流程的办法应对复杂问题（比如 8 步法），但每个环节的误差会累积，最终失败率陡增。对照的是人脑的“潜意识验证机制”：人在行动中不断动态调整路径，而不是机械执行预设流程。Me.bot 的答案是“时间换智能”框架下的内生性训练：白天用 RAG 即时响应，夜间用新数据重训，让记忆和模型能力同步进化。正是这套类脑学习循环（RL’ME’F），让系统能动态适应，而不是靠静态规则撑着。
在幻觉这件事上，Me.bot 的立场刻意反潮流：幻觉是“创造力与风险的共生体”，“有幻觉才会有想象力，才会有发散，才会有智能。”它不做压制，而是用两个机制把幻觉圈住：
- 建立横向监控机制：一套类似人类潜意识的验证程序，持续确认 AI 的行动仍然对准目标；
- 设置参数化约束：通过记忆训练把用户价值观内化为模型的“思维边界”，让幻觉可控，并且为个性化目标服务，而不是拆台。
功能介绍 🍅
“国外会有较多个人用户使用邮件将需要收藏的文件发给自己，国内则更多使用微信传输助手，想要系统地帮个人用户梳理这些原本可能丧失了的记忆。”
Me.bot 的基础功能是把生活里那些离散的点重新连起来，让微信收藏夹里蒙尘的记忆重新派上用场。用户可以输入链接、文件、图片、音频，也可以从 Notion、Evernote、Apple Notes、Markdown 应用和邮箱导入。直接键入的文字若勾选 ‘Todo’，会自动转成日程提醒，并且能准确识别文字里提到的时间，在对应时刻推送。用户与 Me.bot 的历史对话本身，也是它学习的输入数据的一部分。
每存入一条记忆，Me.bot 会思考大约两分钟。思考的产出有两样：highlights 是对这条记忆内容的梳理与总结；memories connected 则把它和历史记忆连起来，试着挖出更深一层的信息，比如用户最近在研究什么、对什么感兴趣。
Me.bot 还会把处理过程变成对话。它向用户抛出一些启发性的问题，比如“浏览过的模型之间的本质区别是什么？”“个性化 AI 将如何改变我们的生活方式？”，引导用户深想一步，甚至带着新问题回来和它讨论。
思考结束后，这条记忆存入 Me.bot 的 Library，并被打上 ‘Tech’、‘Business’、‘Personal Growth’ 之类的标签。除了按时间顺序的流水视图，Library 还提供按话题、类型、来源组织的第二套视图。话题包括 Tech、Moments in Life、Self Growth、Art & Design、Business 等，由 Me.bot 在思考时自动标注；用户也可以自己创建和标注话题，分类体系是活的，不是摊派下来的。类型分类（日程提醒、图片存储、链接存储）按记忆的功能属性划分，找特定类型的信息时特别顺手：想看未来几天的待办，点开 To-dos 就行。
此外还有一个专门的 Discover 板块，每天 9:00 和 19:00 各推送一次。卡片内容紧扣用户自己的输入，多是科技类知识分享和书籍推荐，Me.bot 也会在卡片里发出对话邀请，请用户一起聊聊相关话题。
问它基于最近存储的记忆有什么感悟时，Me.bot 会输出一段总结，回顾近期的阅读和链接。这段总结不止于文本归纳，还带着想象和联想，比如某项技术具体的落地场景。
未来的展望 🍅
陶芳波更远的主张是：个人模型应当成为“AI 身份”，一种人类与世界交互的新型代理主体。他划的界线在模仿与内化之间：AI 身份不是简单模仿用户行为，而是把用户的价值观和决策逻辑内化得足够深，长成一个能自主演化的“数字心识”。当这个身份完全由用户拥有和定义时，技术权力就从中心化平台向个体转移，从通用走向个人，抵达他所说的真正的“AI 平权”。
“我们可以期待的是，未来谈生意，或许可以让你的 me.bot 和我的 me.bot 先聊一聊。”
机器人学有阿西莫夫三定律，陶芳波提出了他的陶式三定律：
- 超级“爱”对齐，即 AI 对人类应该有爱；
- AI 应该跟人有关，体现每个人的不同意志；
- AI 要给人 mindful 的体验。
第三条与其说是原则，不如说是产品立场。陶芳波希望 Me.bot 是一款用起来舒服的产品，而不是像市面上大多数“精神鸦片”那样抢占人的注意力和时间。他的标准很朴素：“让人用完两个小时之后，不会感到后悔在这个 app 上花了两个小时。”他用“干净的水”“清新的空气”“在森林里待着”来描述他想要的 mindful。
当被问到 AI-native Memory 打造的“Second Me”能不能通向数字永生时，陶芳波自称“人类中心主义者”，给出了一个这个赛道上少有创始人会给的回答：世界上最伟大的发明就是死亡。人的生命凋零之后，数字永生会带来很多伦理问题，我们的社会还没有做好准备。他希望避开数字永生，不与死亡对抗。
“人生本身是一个游乐场，首先你要确定好你想玩的游戏，第二呢是怀着爱意去玩这个游戏。我觉得这是最重要的东西。”
这个回答里藏着整个项目的底色：AI 不该替代人，而该成为人的延伸。
Preface: When memories evolve from indexed fragments into a reasoning-capable, evolvable parameterized network through temporal training, the core mission of knowledge engineering shifts from “What is true?” to “What is ME?”, marking the tipping point where knowledge graphs transcend universal cognitive infrastructure to embrace personalized mind science. This exploration is fundamentally an experiment in human cognition, and our research on entity relationships and dynamic graphs may hold the key to unlocking it.
Me.bot leverages personalized individual data rather than generalized human data. In this emerging social paradigm, before you and I meet in person, let our bots, embodying our unique cognitive spirits, converse first.
Vision Behind the Product 🍅
Tao Fangbo, founder of Mindverse, introduced their product Me.bot - Your AI Second Brain in an interview:
“Me.bot goes beyond RAG or long-context knowledge retrieval. It trains personalized models based on user memories to deliver proactive, individualized insights. Our goal is to step back from the world and step closer to the individual.”
While ChatGPT and Deepseek compete over general-purpose intelligence, Mindverse is betting on the opposite axis: a dedicated “me” for every user, not a lifestyle assistant, but a digital extension of the self. The bet compresses into one sentence: rather than predicting the next token, predict your next thought.
The claim is testable, and Tao offers an anecdote as evidence. Mindverse trained his personal Me.bot, and on his 35th birthday it generated a statement nearly identical to his own words: “I’m Tao Fangbo, founder of Mindverse. Today, as I reflect on my 35th birthday, I’m deeply grateful for the decision I made at 30; it was unequivocally right.”
Asked how this is achieved, Tao’s answer was one word: “Memory.” Human memories are the ultimate dataset (dense, longitudinal, and unique to one person), and they make it possible to predict the flow of an individual’s cognitive river.
Me.bot and Shared Memory 🍅
The architectural distinction matters here. Traditional AI memory systems (RAG, long context) treat memory as external data to be retrieved: the model stays generic, and personalization lives in what gets fetched into the prompt. Me.bot instead parameterizes user memories, internalizing lived experience and cognitive patterns into the model’s weights. The design echoes neuroscience’s axiom that “memory shapes identity”: once an AI can predict the “next thought” from temporal life data (what you hear, see, and contemplate), it stops being a reactive tool and becomes a cognitive digital twin.
The key differentiator is the temporal training framework. Where large language models freeze their parameters for months at a time, Me.bot emulates human growth: just as the brain consolidates memories during sleep, Me.bot compresses each day’s experiences into updated parameters nightly. Tao frames the alignment consequence sharply: “RLHF (Reinforcement Learning from Human Feedback) aligns AI with humanity’s collective interests. But my personal model must align with my individual interests; this is the shift from RLHF to RL’ME’F.”
Tao is equally direct about what he sees as the limits of mainstream agent frameworks (Coze, Workflow-style systems): they are “scaffolding external logic,” assembling intelligence out of externally wired steps. An agent that answers a complex request by decomposing it into a fixed pipeline (an eight-step method, say) accumulates error at every step, and the failure rate compounds. Contrast that with the brain’s “subconscious validation mechanism”: humans continuously adjust their path mid-action rather than mechanically executing a preset plan. Me.bot’s answer is “endogenous training” under a “time-for-intelligence” framework: serve requests during the day via RAG, retrain on the day’s new data at night, so memory and model capability evolve in sync. The brain-like learning loop (RL’ME’F) is what lets the system adapt dynamically instead of leaning on static rules.
On hallucination, Me.bot takes a deliberately contrarian position: hallucination is “a symbiosis of creativity and risk”: “only with hallucinations can there be imagination, divergence, and intelligence.” Rather than suppressing it, the design constrains it with two mechanisms:
- A lateral monitoring mechanism, a subconscious-like validation program that continuously checks whether the AI’s actions still align with its goals;
- Parameterized constraints, internalizing the user’s values into the model as “thought boundaries,” so hallucination stays controllable and serves personalized goals rather than undermining them.
Features 🍅
“While global users email themselves files, Chinese users rely on WeChat file transfer assistants. Me.bot systematically rescues these drowning memories.”
Me.bot’s core function is connecting life’s scattered dots, reviving the artifacts buried in WeChat collections and similar dumping grounds. Users can input links, files, images, and audio, or import from Notion, Evernote, Apple Notes, Markdown apps, and email. Text entries tagged ‘Todo’ auto-convert to smart reminders, with timestamps parsed for precise alerts. The user’s conversation history with Me.bot itself also feeds back into the training data.
When a memory is added, Me.bot processes it for roughly two minutes. The output of that thinking pass is twofold: highlights, which summarize and organize the memory’s content, and connected memories, which link it to historical entries in search of deeper patterns: what is the user currently studying? What are they drawn to?
Me.bot also turns processing into dialogue. It poses thought-provoking questions back to the user, such as “What is the fundamental difference between the models you have browsed?” or “How will personalized AI change our way of life?”, designed to pull the user into reflection and surface new questions worth discussing.
Once processed, the memory lands in Me.bot’s Library, tagged with categories such as ‘Tech’, ‘Business’, or ‘Personal Growth’. Beyond the chronological stream, the Library offers a second organization by topics, types, and sources. Topics include Tech, Moments in Life, Self Growth, Art & Design, Business, and so on, auto-tagged during processing, but users can also define and apply their own topics, which keeps the taxonomy flexible rather than imposed. Type categories (schedule reminders, image storage, link storage) group memories by functional property, which pays off when hunting for a specific kind of information: checking upcoming tasks means simply opening the To-dos category.
There is also a dedicated Discover section, with two daily pushes at 9:00 and 19:00. The cards are grounded in the user’s own inputs (knowledge sharing in technology, book recommendations), and Me.bot invites the user to discuss them.
Asked for insights over recently stored memories, Me.bot produces a review of recent readings and links that goes beyond textual summary, adding imagination and association, for instance concrete application scenarios for a technology the user has been reading about.
Future Outlook 🍅
Tao’s long-range claim is that the personal model should become an “AI identity,” a new type of agent through which humans interact with the world. The distinction he draws is between mimicry and internalization: the AI identity does not simply imitate user behavior; it absorbs the user’s values and decision logic deeply enough to form a self-evolving “digital mind.” And when that identity is fully owned and defined by the user, technological power shifts from centralized platforms to individuals, from general-purpose to personalized, toward what he calls true “AI equality.”
“We can expect that in the future, when discussing business, perhaps your me.bot and my me.bot can chat first.”
Alongside Asimov’s Three Laws of Robotics, Tao proposes his own “Tao’s Three Laws”:
- Super “love” alignment (i.e., AI should have love for humans);
- AI should be relevant to humans, reflecting each individual’s will;
- AI should provide a mindful experience.
The third law is a product stance as much as a principle. Tao wants Me.bot to feel comfortable to use, unlike the “mental opiates” that dominate the market by seizing attention and time. His bar is simple: “something that doesn’t make you regret spending two hours on the app after using it.” He likens the mindfulness he is after to “clean water,” “fresh air,” or “being in the forest.”
Asked whether an AI-native memory-built “second me” could enable digital immortality, Tao identified himself as “human-centric” and gave an answer few founders in this space would: the greatest invention in the world is death. Digital immortality after a human life fades raises ethical questions our society is not prepared for, and he intends to avoid it rather than fight death.
“Life itself is an amusement park; first, you need to figure out what game you want to play, and second, play it with love. I think that’s the most important thing.”
The answer distills the project’s core conviction: AI should not replace humans, but extend them.
