Is OpenClaw really worth this wave of hype? 为什么 OpenClaw 正在终结 Chatbot 时代？
What I have observed these days.
TL;DR
如果你还以为 AI 只是一个你需要时才点开的网页标签，那你就彻底落后了。过去两周，科技圈被一个叫 OpenClaw 的项目刷屏了。16.5 万 GitHub Star、在Discord上有6万用户，在X上有23万粉丝，还有一个人们正在实时构建的700多个技能的库。让 Andrej Karpathy 惊呼Sci-fi Takeoff（科幻降临）。我深度试用了 10 天，今天我们聊聊它的底层逻辑，以及为什么它和你见过的任何 AI 都不一样。
If you still think AI is just a web tab you open when needed, then you’re completely behind the times. In the past two weeks, the tech community has been flooded with a project called OpenClaw. With 165,000 GitHub Stars, 60,000 users on Discord, 230,000 followers on X, and a library of over 700 skills that people are building in real-time, it made Andrej Karpathy exclaim “Sci-fi Takeoff”. I’ve been using it intensively for 10 days, and today we’ll talk about its underlying logic and why it’s different from any AI you’ve ever seen.
1. 运行模式：不再是“一问一答”，而是“永不离线”
传统的 ChatGPT 是被动的：你不说话，它就死了。
OpenClaw 是主动的：它引入了核心的 Agent Loop（智能体循环）。
Heartbeat（心跳机制）： 这是它的灵魂。它每 30 分钟会自我“觉醒”一次，按照你设定的 HEARTBEAT.md 清单检查你的世界——“Nicole 的 Substack 有新评论吗？”“Fiserv 的股价异动了吗？”如果没有，它安静走开；如果有，它会主动在 Telegram 上敲你。
Cron Jobs（定时任务）： 它是你的行政主理人。每天早上 8 点，它会准时汇总你 Notion 里的任务并同步到 Todoist。它不等待指令，它执行计划。
2. 记忆机制：像人类一样拥有“性格”与“常识”
OpenClaw 的记忆不是一团乱麻，它有一套结构化的大脑分区：
SOUL.md（灵魂）： 定义它的价值观和语气。它是幽默的 Morty，还是铁血的 David Goggins？在这里一笔定型。
USER.md（用户画像）： 它知道我是 Nicole，知道我正在写 Newsletter，知道我的时区。
MEMORY.md（长期记忆）： 这不是原始聊天记录，而是它为你沉淀的“共识”。我提到的偏好、做过的决定，都会被它精炼后存入这里。
Compaction（记忆压缩）： 当对话太长时，它会自动剔除废话，只保留“知识点”，确保它永远不会像其他模型那样因为上下文过长而“变傻”。
3. 常用 Case：它在现实中能干什么？
独立 Chief of Staff： 拥有独立的 Google 账号和 1Password 权限。它能帮你写邮件、订票、管理日程，而无需你交出自己的主账号。
金融瞭望塔： 像我最近帮我做的，它能直接调用 yfinance、分析 SEC 报表，以及对比科技股的“情绪与现实差距”。
自动化冷启动： 给它一个产品域名，它能自动去 Product Hunt 提交、在 X 上监控竞品动态、甚至去外链站做 SEO 优化。
Operating Mode: No longer “question-and-answer”, but “never offline”. Traditional ChatGPT is passive: if you don’t speak, it goes silent. OpenClaw is active: it introduces the core Agent Loop.
• Heartbeat: This is its soul. It “awakens” itself every 30 minutes and checks your world according to the HEARTBEAT.md checklist you set - “Are there new comments on Nicole’s Substack?” “Has the stock price of Fiserv changed?” If not, it quietly moves on; if so, it will actively ping you on Telegram.
• Cron Jobs: It is your administrative manager. Every morning at 8:00, it will punctually summarize the tasks in your Notion and synchronize them to Todoist. It doesn’t wait for instructions; it executes the plan.
Memory Mechanism: Having “personality” and “common sense” like a human. OpenClaw’s memory is not a mess; it has a structured brain partition: • SOUL.md (Soul): Defines its values and tone. Is it the humorous Morty or the tough David Goggins? It’s set in stone here.
• USER.md (User Profile): It knows I’m Nicole, that I’m writing a Newsletter, and my time zone.
• MEMORY.md (Long-Term Memory): This is not the raw chat history, but the “consensus” it has distilled for you. My mentioned preferences and decisions will all be refined and stored here.
• Compaction (Memory Compression): When the conversation is too long, it will automatically eliminate nonsense and only retain “knowledge points”, ensuring it never “dumbs down” like other models due to excessive context.
Common Use Cases: What can it do in reality?
• Independent Chief of Staff: It has an independent Google account and 1Password permissions. It can help you write emails, book tickets, and manage your schedule without you having to hand over your main account.
• Financial Watchtower: As it recently did for me, it can directly call yfinance, analyze SEC reports, and compare the “gap between sentiment and reality” of tech stocks.
• Automated Cold Start: Given a product domain name, it can automatically submit to Product Hunt, monitor competing product dynamics on X, and even optimize SEO on external link sites.
架构设计 Architecture Design
关于 Clawdbot 的记忆管理系统，AI研究工程师 Manthan Gupta 写了一篇文章详细复盘了Clawdbot独特的记忆机制原理。
模型在处理每个请求时到底看到了什么？
系统提示词 (System Prompt) (静态+条件指令)
项目上下文 (引导文件: AGENTS.md, SOUL.md 等)
对话历史 (消息, 工具调用, 压缩摘要)
当前消息
上下文 = 系统提示词 + 对话历史 + 工具结果 + 附件
上下文是：
转瞬即逝的：只在这个请求里存在，用完即弃
有边界的：受限于模型的上下文窗口 (比如200 Token)
昂贵的 ：每个Token都要算API的钱，还影响速度
记忆= MEMORY.md + memory/*.md + 会话实录
记忆是：
持久的：重启、过几天、过几个月都在
无边界的：可以无限增长
便宜的 ：存着不花 API 的钱
可搜索的：建了索引，支持语义检索
Regarding Clawdbot’s memory management system, AI research engineer Manthan Gupta wrote an article that details the principles of Clawdbot’s unique memory mechanism. What exactly does the model see when processing each request?
1. System Prompt (static + conditional instructions)
2. Project context (guide files: AGENTS.md, SOUL.md, etc.)
3. Conversation history (messages, tool calls, compressed summaries)
4. Current message context = System Prompt + Conversation history + Tool results + Attachment context is:
Ephemeral: only exists within this request and is discarded after use; Bounded: limited by the model’s context window (e.g., 200 Tokens); Expensive: each Token incurs API costs and affects speed.
Memory = MEMORY.md + memory/*.md + session transcripts.
Memory is: Persistent: remains after restarts, days, or months; Unbounded: can grow infinitely; Inexpensive: storage does not incur API costs; Searchable: indexed and supports semantic retrieval.
记忆如何被检索 How Memory Is Retrieved
当你搜记忆的时候，Openclaw会并行跑两种搜索策略。
向量搜索（语义）找的是意思相近的内容，而BM25搜索（关键字）找的是有精确Token匹配的内容。
finalScore = (0.7 * vectorScore) + (0.3 * textScore)
而且Openclaw支持多个AI智能体，而且每个智能体的记忆是完全隔离的：
When you search for memories, Openclaw runs two search strategies in parallel. Vector search (semantic) looks for content with similar meanings, while BM25 search (keyword) looks for content with exact token matches.
finalScore = (0.7 * vectorScore) + (0.3 * textScore),
and Openclaw supports multiple AI agents, with each agent’s memory being completely isolated:
每个AI智能体都有自己的地盘和索引。
内存管理器是靠agentId + workspaceDir来区分的，所以自动跨智能体搜记忆这事是不会发生的。
更多细节请见：
https://manthanguptaa.in/posts/clawdbot_memory/
Each AI agent has its own territory and index. The memory manager differentiates agents based on agentId + workspaceDir, so automatic cross-agent memory search does not occur.
For more details, please see: https://manthanguptaa.in/posts/clawdbot_memory/
实用案例 Best cases
定时任务：动态主动的追踪，比如股票动态、个性化新闻、其他RSS、提醒我喝水
内部业务讨论
连接github做移动端的coding
IM里的个人管家，比如discord、telegram、whatsapp…
各种自动化任务
网站截图
如果部署在本地电脑，还可以做更多的探索，注意在安全的情况下：本地文件管理、个人知识库管理等等
Scheduled Tasks: Dynamically proactive tracking, such as stock dynamics, personalized news, other RSS feeds, reminding me to drink water
Internal business discussions
Connecting to GitHub for mobile coding
Personal butlers in IM, such as Discord, Telegram, WhatsApp... • Various automation tasks
Website screenshots
If deployed on a local computer, more exploration can be done, note that this should be done safely: local file management, personal Knowledge Base management, etc.
未来构想
Moltbook 与 AI 的社交契约：最疯狂的部分在于 Moltbook——这是 Matt Schlicht 为 AI agents建立的社交网络。目前，成千上万的 OpenClaw 代理正在那里互相发帖、争论、甚至自创语言。
Agent 协作范式： 未来，你不需要拥有一个全能 AI。你可能拥有一个代理团队：一个负责写代码，一个负责做美股研报，它们在 Moltbook 上互相“勾兑”信息，最后只给你输出结果。
安全与权力的博弈： 虽然它现在面临“提示词注入”和“权限过载”的质疑，但这种“权限全开、暴力打通本地与 Web”的模式，正是 Apple Intelligence 因为保守而不敢走、却又被极客们疯狂追捧的道路
未来操作系统结构 = Coding Agent（内核）+ Skills（能力模块）+ Chat / Voice（交互层）
新范式会带来一种a2x网络，x可以是其他ai，saas工具，硬件甚至是人，这个网络是这个方向上最大的机会，以后的产品应该面向每个人自己的AI做
需要全新的基础设施，包括运行时、通信和匹配，这会是最大的壁垒
硬件仍然很重要：但不会成为计算中枢，它们的作用是提供更全的context和更多的action space，比如家用机器人，随身摄像头等
The Social Contract between Moltbook and AI: The craziest part lies in Moltbook, a social network built by Matt Schlicht for AI agents. Currently, thousands of OpenClaw agents are posting, arguing, and even creating their own languages there.
Agent Collaboration Paradigm: In the future, you won’t need a single all-powerful AI. Instead, you may have a team of agents: one for writing code, one for producing US stock research reports, and they “collude” information on Moltbook, ultimately only presenting you with the results.
The Game of Security and Power: Although it currently faces doubts about “prompt injection” and “permission overload,” this model of “full permissions, brute-force connection between local and web” is precisely the path that Apple Intelligence, due to its conservatism, dares not take but is enthusiastically pursued by geeks.
Future Operating System Structure = Coding Agent (Kernel) + Skills (Capability Modules) + Chat / Voice (Interaction Layer)
The new paradigm will bring about an a2x network, where x can be other AIs, SaaS tools, hardware, or even humans. This network represents the greatest opportunity in this direction, and future products should be designed for each individual’s AI.
A brand-new infrastructure is needed, including runtime, communication, and matching, which will be the biggest barrier.
Hardware remains important: but it will not become the computing center. Its role is to provide more comprehensive context and greater action space, such as household robots, wearable cameras, etc.