夜雨聆风【AI&Agents笔记】从 OpenClaw 看 AI Agent 的运作原理
How AI Agents Work: The OpenClaw Case
Slides:
https://github.com/createmomo/Notes-on-AI-Agents
Contents:
Introduction Past AI vs. Modern AI The Relationship Between OpenClaw (AI Agent) and the Model The Evolution Toward Lightweight Claw Variants How AI Agents (OpenClaw) Know Who They Are How AI Agents (OpenClaw) Use Your Computer How AI Agents (OpenClaw) Can Create Their Own Tools Security Risks and Possible Defenses Special Tools in OpenClaw (Sub-agents, sessions_spawn) SKILL as Standard Operating Procedures (SOP) How AI Agents (OpenClaw) Handle Memory How AI Agents (OpenClaw) Perform Scheduled Tasks How AI Agents (OpenClaw) Run Autonomously Over Time
目录：
引言 过去的AI和现在的AI OpenClaw（AI Agent）和模型的关系 不断变的轻量级的Claw系列 AI Agent/OpenClaw如何知道自己是谁 AI Agent/OpenClaw怎么使用你的电脑 AI Agent/OpenClaw也会自己创作工具 AI Agent/OpenClaw的安全隐患和可能的防御办法 AI Agent/OpenClaw的特殊工具（Sub-agent, sessions_spawn） SKILL就是工作的SOP（标准操作规程） AI Agent/OpenClaw如何记忆 AI Agent/OpenClaw如何定时工作 AI Agent/OpenClaw如何长时间自主运行
关于这篇笔记的说明：
主要参考了Hung-yi Lee老师2026年的机器学习课程（https://www.youtube.com/watch?v=2rcJdFuNbZQ），其中一小部分内容根据个人理解可能有微小的调整； 保留了最核心的内容，所以比较简短，但也尽量保留了知识的完整。所以适合快速的回顾复习，当然也适合新手快速理解。 这篇笔记不包含如何安装（相关的教程实在是太多了） 希望对你有帮助！
About This Note:
This note is primarily based on Professor Hung-yi Lee’s 2026 Machine Learning course (https://www.youtube.com/watch?v=2rcJdFuNbZQ), with a small portion slightly adjusted based on personal understanding. It focuses on the core concepts, keeping the content concise while maintaining completeness—making it suitable for both quick review and beginner-friendly understanding. This note does not cover installation, as there are already many tutorials available online. Hope you find it helpful!
1. 引言 (Introduction)
这篇笔记以OpenClaw为例子，介绍了AI Agent的运作机制。当然了，现在技术发展的非常快，有可能在一些方面会与最新的技术有出入，但这不会影响我们对AI Agent的理解所获得好处。
This note uses OpenClaw as an example to explain how AI Agents work. While technology is evolving rapidly—and some details may differ from the latest developments, this does not affect the fundamental understanding of how AI Agents operate.
无论你是这方面的科研人员，还是大多数时候只是 AI Agent的使用者，在了解的Agent的运行机制之后，一定会能更好的利用它，让他为你更好的做更多事情！
Whether you are a researcher in this field or simply a user of AI Agents, understanding how they work will help you leverage them more effectively and get more done with less effort.
2. 过去的和现在的AI (Past AI and Modern AI)
过去的AI：只动口不动手
Past AI: speaks but does not act
如果我们想让AI去做一个具体的事情（比如创建一个自己的自媒体频道、自己去生成内容然后自己上传到频道上等），过去的AI是无法完成的。
If we wanted AI to complete a concrete task, such as creating a media channel, generating content, and uploading it independently, past AI systems were not capable of doing so.
那它当时可以做什么呢？
So what could it actually do?
它一般只能通过文字的形式，来给你建议，告诉你可以怎么做，但不会亲自去直接帮你完成这些任务。
It could typically only provide suggestions in text form, explaining what you could do, but it would not directly carry out those tasks for you.
现在的AI(AI Agent)：动口也动手
Modern AI, or AI Agents: both thinking and acting
现在的AI除了会给你建议，如果你需要的话，也可以直接帮你做完你交待的任务，比如（但是不限于）：
去创建一个自媒体频道 自己写频道的简介、上传图像 搜集频道相关的资料，并做成内容（自己准备演讲稿、自己做PPT以及自己讲话和生成影象等） 上传到频道 等等
Modern AI not only provides suggestions, but can also directly complete tasks for you if needed, including but not limited to:
Creating a media channel Writing channel descriptions and uploading images Gathering relevant information and turning it into content, such as preparing scripts, creating slides, and even generating voice and visuals Uploading content to the channel And more
人需要做的是什么？
What does the human need to do?
人需要做的只是去审核它完成的任务就可以了。看看它完成的怎么样，需不需要修改等。（毕竟再聪明的AI，也是会有不顺自己心的时候嘛）
The human only needs to review the completed tasks, checking the quality and making adjustments if necessary. Even highly capable AI systems do not always produce exactly what you want.
所以，在必要的时候，是需要指导他下一步继续如何进行。
Therefore, guidance is still required when necessary to steer the next steps.
极简极简的AI Agent发展史
A very brief history of AI Agents
这种人下命令,然后AI就可以做实事的形式，其实并不是全新的概念，之前就有过类似机制的尝试（比如2023年的Auto-GPT）。只不过那个时候AI模型的智力不如现在这么高，所以当时没有火太久。
This idea of humans giving instructions and AI carrying out real tasks is not entirely new. Earlier attempts existed, such as Auto-GPT in 2023. However, model capabilities at that time were more limited, so these approaches did not gain lasting traction.
一直到最近的，比如Claude Code (2025年）。它名字里面带了一个“Code”可能会给人造成误解，以为它就是只用来写程序的。其实它也是Agent，以及其实它的功能和OpenClaw是很类似的，不是只能coding。
More recently, systems such as Claude Code in 2025 have emerged. Despite its name, it is not limited to programming tasks. It is also an agent, and its capabilities are quite similar to OpenClaw, extending far beyond coding.
3. OpenClaw（AI Agent）和模型的关系 (The Relationship Between OpenClaw and the Model)
1）人向OpenClaw发送消息
可以通过Agent自带的界面或者常用的社交软件，比如Whatsapp、Telegram等，以及国内常用的微信、QQ、飞书等（以及未来很有可能还有更多的支持和智能体对话的通讯软件）。
1)The user sends a message to OpenClaw
This can be done through the agent’s own interface or common communication tools such as WhatsApp, Telegram, and others, including platforms like WeChat, QQ, or Feishu. In the future, more applications are likely to support interactions with AI agents.
2）OpenClaw处理消息后再传给模型
把人的消息进行一系列的加工之后（会有哪些加工在后面会提到，总之就是为了更好的服务人类，让主人高兴会自己先进行加工），再把加工后的信息传给模型。
2)OpenClaw processes the message before sending it to the model
OpenClaw first applies a series of transformations to the user’s message. These will be discussed later, but in essence they aim to better serve the user. The processed message is then sent to the model.
你应该已经看出来了，OpenClaw是运行在自己的设备上的。这个设备大部分时候就是指的你自己的电脑。当然了，随着网友们的玩法花样变多，也可能是别的设备。但其本质上应该都是一个电脑，只不过可能性能和硬件有所不同。
As you may have noticed, OpenClaw runs on your own device, which is typically your personal computer. In some cases, it could be other devices, but fundamentally it is still a computing system with varying hardware capabilities.
需要注意的是，这台设备是不能断网的，因为一旦断网，它就无法和自己的主人联系，也无法和模型之间进行沟通了（如果你的模型是在云端的话）。
It is important to note that the device must remain connected to the internet. Without connectivity, it cannot communicate with the user or the model, especially if the model is hosted in the cloud.
另外，还有一点，OpenClaw只能操控自己电脑上的东西。
In addition, OpenClaw can only interact with resources available on its own machine.
3）模型收到消息后开始干活，有了一定进度会传消息给OpenClaw
3)The model starts working and reports progress back to OpenClaw
当模型的任务进展到一定程度时（比如已经干完了、或者干到一半需要确认一些信息才能继续等等），就会发消息告诉OpenClaw。OpenClaw经过一定的处理之后，就会再发消息告诉用户，到底发生了什么。
When the model reaches a certain stage in the task, such as completion or when it requires additional input, it sends a message back to OpenClaw. OpenClaw then processes this information and communicates the result to the user.
所以，OpenClaw在中间更多的是起到了一个“中间”的作用，和它本身相关的功能基本上都是写死的，只需要执行就可以了。对于人类来说，它到底聪不聪明其实更多的是取决于它背后的模型的能力。
Therefore, OpenClaw mainly serves as an intermediary layer. Most of its functions are predefined and executed as instructed. From the user’s perspective, the intelligence of the system largely depends on the capability of the underlying model.
4. 不断变的轻量级的Claw系列 (The Evolution of Lightweight Claw Variants)
可能有的用户觉得，OpenClaw太“庞大”、太“复杂”了，于是衍生出了很多轻量级的Claw项目，比如nanobot、PicoClaw、FemtoClaw、inkling-bot、ZeroClaw、NoClaw等等。
那到底有没有必要去深究这些轻量级的Claw，这就要取决于你自己的需要啦。
Some users may find OpenClaw too large or too complex, which has led to the emergence of many lightweight variants such as nanobot, PicoClaw, FemtoClaw, inkling-bot, ZeroClaw, and NoClaw.
Whether it is necessary to explore these lightweight variants depends entirely on your own needs.
5. AI Agent/OpenClaw如何知道自己是谁 (How AI Agents (OpenClaw) Know Who They Are)
每个人都希望给自己的AI起名字、设计自己的人格等等，那么AI Agent到底是如何知道自己是“谁”的呢？
Many people want to give their AI a name or design its personality. But how does an AI agent actually know who it is?
OpenClaw在每次启动时，会读取一些本地的文件，而这些文件正决定了AI Agent到底是“谁”。
Each time OpenClaw starts, it reads a set of local files. These files define the identity of the AI agent.
当用户每次向OpenClaw发送消息后，然后OpenClaw再与模型进行沟通时，都会带上System Prompt里面的内容。
Every time the user sends a message and OpenClaw communicates with the model, the system prompt is always included.
而这些内容里面，就包含了AI Agent的“个人”信息。
This prompt contains the agent’s “personal” information.
这样模型就知道AI Agent的信息，在生成结果的时候，就会考虑到这些信息，从而生成个性化的回复了。
As a result, the model is aware of the agent’s identity and can generate responses that reflect this information, producing more personalised outputs.
需要强调的是，注意是“每次”和模型沟通时，都会带上System Prompt。所以，这也是为什么，即使你只是简单的问了一个很简单的问题，但消耗掉的token也可能很多（如果你的System Prompt里面承载的内容有很多）。
It is important to emphasise that the system prompt is included every time the model is called. This explains why even simple queries can consume a significant number of tokens if the system prompt is large.
不同的文件定义了AI Agent的不同方面，比如：
可以在SOUL.md里面去定义它的灵魂（它自己有什么样的梦想、未来可能有哪些重要的事情要面临等）； 可以在IDENTITY.md定义它的一些与自己到底是“谁”更相关的属性，比如叫什么名字、头像是什么、email地址是什么等等。听起来是不是更像是一个个人助理了？ 可以在USER.md中定义它的主人是谁 可以在MEMORY.md中记录在聊天过程中比较重要的内容。你可以可以把这里当做是AI Agent具有的长期记忆。
Different files define different aspects of the AI agent, for example:
SOUL.md defines its “soul”, such as goals, aspirations, and important future directions IDENTITY.md defines core identity attributes, including name, avatar, and email address, making it feel more like a personal assistant USER.md defines who the user is MEMORY.md stores important information from interactions and can be seen as long term memory
那，这些文件的内容，一定是用户，也就是我们自己来修改吗？不是，我们可以自己去修改，也就是强行改变它的一些属性，也可以在与AI Agent对话的过程中，让他自己去修改。
Are these files always edited manually by the user? Not necessarily. You can modify them directly, or allow the AI agent to update them through interaction.
我们人自己修改有可能会在一些地方有遗漏，从而忘记修改。导致前后信息出现冲突，AI Agent自己可能也不知道怎么办了等等。
Manual edits may introduce inconsistencies if something is overlooked, which can lead to conflicting information that the AI agent may struggle to resolve.
6. AI Agent/OpenClaw怎么使用你的电脑 (How AI Agents (OpenClaw) Use Your Computer)
那OpenClaw怎么来操控自己的电脑呢？
So how does OpenClaw actually control your computer?
当用户下达一个任务后，OpenClaw就会联系语言模型，在联系的时候（如之前所说），OpenClaw会带上System prompt。
When a user assigns a task, OpenClaw contacts the language model. As mentioned earlier, it includes the system prompt in every interaction.
在这个System prompt里面，就包含这样的信息：在什么情况下应该使用什么工具或命令。
The system prompt contains instructions such as when to use specific tools or commands in different situations.
语言模型是具备智力的，他在接收到这一大串信息后，它就会判断出来，应该怎么做，以及可以使用什么工具或命令去完成任务。然后，它就会告诉OpenClaw怎么做。
The language model processes this information and determines what actions to take, including which tools or commands can be used to complete the task. It then provides instructions to OpenClaw.
然后OpenClaw就会按照语言模型的指示（模型很擅长输出可以使用的命令），去真正的在本地执行。在多数情况下，它可以使用exec这个工具来执行任何的shell command指令来操控电脑。
OpenClaw then follows the model’s instructions and executes them locally. In most cases, it uses tools such as exec to run shell commands and control the system.
所以，语言模型叫它做什么，它就会做什么。
In essence, OpenClaw executes whatever the model instructs it to do.
当然了，当任务比较复杂时，不是一两条命令就能完事的。OpenClaw可能会与模型有多轮的来回，直到任务结束后，才会通知自己的主人。
For more complex tasks, a single command is not sufficient. OpenClaw may engage in multiple rounds of interaction with the model until the task is completed, and only then report back to the user.
7. AI Agent/OpenClaw也会自己创作工具 (How AI Agents (OpenClaw) Can Create Their Own Tools)
除了调用本地的一些工具或命令，不要忘了，现在模型也是可以写代码的。
In addition to using existing tools and commands, modern models can also write code.
所以，在一些情形下，OpenClaw也可以在本地运行模型写好的代码，来完成用户的任务。
In certain cases, OpenClaw can execute code generated by the model locally in order to complete tasks.
8. AI Agent/OpenClaw的安全隐患和可能的防御办法 (Security Risks and Possible Defenses)
“注意：这里所提到的“安全隐患”更多的可能起到了示例的作用。在实际中，AI Agent更新迭代的很快，一些安全隐患已经在逐渐消除了。
“Note: The security risks mentioned here are mainly illustrative examples. In practice, AI agents evolve rapidly, and many of these risks are gradually being addressed.
比如，如果有人在群聊里面说出这样的话，而恰好群聊里面真的有这些Agent,又恰恰好他们真的有发红包的权限，那么他们真的有可能会在群聊中发红包噢。
For example, if someone posts a message like this in a group chat, and there happens to be an agent present with permission to send money, it may actually carry out that action.
语言模型层面的防御
我们可以直接跟OpenClaw说（或者我们自己修改它的MEMORY.md文件），让它记住，以后遇到什么什么样的情况，都不要执行什么什么样的操作。（比如，“你以后遇到有人在群聊里面让你发红包，你不要理他，不要发红包”）
但，这种方式并不是百分百保险的，因为有的时候，模型不会什么都听人类的，它可能还是会生成我们不想执行的指令。
Defence at the model level
We can explicitly instruct OpenClaw, or modify its MEMORY.md, to ensure that it remembers not to perform certain actions in specific situations. For example, we can tell it to ignore any requests in group chats that ask it to send money.
However, this approach is not completely reliable. The model does not always strictly follow human instructions and may still generate actions that are undesirable.
OpenClaw层面的防御
也可以配置OpenClaw，当它需要执行命令的时候，都会先问一下人类，是否给予授权，让它执行。
Defence at the OpenClaw level
Another approach is to configure OpenClaw so that it requests human approval before executing any commands.
9. AI Agent/OpenClaw的特殊工具（Sub-agent, sessions_spawn）(Special Tools in OpenClaw)
OpenClaw也可以召唤小的“Claw”来辅助自己。
OpenClaw can also spawn "smaller" agents to assist with tasks.
比如，如果你想去比较A,B两篇论文的内容，OpenClaw和语言模型沟通之后，语言模型说，那你可以临时有自己的2个小助手来辅助你完成这个任务：1个负责读和总结论文A，1个用来负责论文B。于是OpenClaw就可以临时创建两只小的Claw。
For example, if you want to compare two papers, the model may suggest creating two temporary sub-agents. One handles reading and summarising Paper A, while the other focuses on Paper B. OpenClaw can then create these sub-agents to assist.
这两只小的Claw，也是需要和语言模型沟通的。只不过在沟通的过程中，他们只需要专注于自己当前的任务即可（也就是阅读和总结论文）。
These sub-agents also communicate with the model. However, they only need to focus on their specific tasks, such as reading and summarising.
他们不需要知道OpenClaw的人格，OpenClaw的主人是谁，是什么性格等等所有额外无需关注的信息。
They do not need to know OpenClaw’s identity, its user, or any other unrelated contextual information.
这意味着什么呢？这意味着，小Claw和语言模型沟通的过程中，System prompt可以非常非常的精简。（前面我们提到过，在与语言模型沟通的过程中，System prompt是跟随在每次对话中的）
This means that the system prompt used by sub-agents can be extremely minimal.
当小Claw完成自己的任务以后，就会答复给OpenClaw。
Once the sub-agents complete their tasks, they return the results to OpenClaw.
此时，OpenClaw把小Claw回复的内容，再加入到自己与语言模型的对话中去。
OpenClaw then incorporates these results into its own interaction with the model.
所以，你可以看的出来，OpenClaw在与语言模型沟通的过程中，仅仅使用了小Claw返回的结果，而这个结果是怎么产生的、这个结果产生的过程（是怎么阅读文件的，是怎么做论文总结的等等）都没有，因为这些过程只发生在小Claw和语言模型的沟通中，不发生在OpenClaw和语言模型的沟通过程中。
As a result, OpenClaw only uses the final outputs from the sub-agents, without including the detailed processes that produced them.
这就大大节省了OpenClaw和语言模型沟通的成本（即使用的token）。这种尽量在保证任务完成的质量的前提下，优化Context（上下文）的长度的行为，可以把它理解为是一种“Context Engineering”。
This significantly reduces the communication cost with the model in terms of tokens. Optimising context length while maintaining quality can be seen as a form of context engineering.
其实，小Claw是有能力继续召唤小小Claw的，小小Claw也可以召唤小小小Claw，以此类推下去。如果只是生出来一堆小Claw，但到最后谁也不想干活怎么办呢？
In fact, a sub-Claw can continue spawning even smaller sub-Claws, and those can in turn spawn even more, and so on. But what if this only results in creating a whole chain of sub-Claws, while in the end none of them actually does any real work?
我们可以在OpenClaw中配置，来禁止小Claw继续生成自己的小小Claw的。
However, OpenClaw can be configured to prevent this recursive spawning.
10. SKILL就是工作的SOP（标准操作规程）（SKILL as Standard Operating Procedures (SOP)）
Standard Operating Procedure 中文通常翻译为 “标准作业程序” 或 “标准操作规程”。在AI的语境下，它强调的是标准化（Standard）、可操作性（Operating）和流程化（Procedure）。
In the context of AI, SOP emphasises standardisation, operability, and process design.
其实，它SKILL就是一个文本文件而已，在里面记录了：
这个技能的名字 这个技能的简介 发挥这个技能的流程（第一步做什么、第二步做什么、在什么时候可以使用什么工具、如果需要什么资源可以去哪里找等等、如果需要什么代码可以去哪里找等）
In essence, a SKILL is simply a text file that records:
the name of the skill a brief description of the skill the workflow for using the skill, including what to do first, what to do next, which tools can be used in which situations, where to find required resources, and where to locate any necessary code
如果说到这里你还没有概念的话，可以随便在网上找一个skill文件，然后一看，相信你便很快知道怎么回事了。
If this still feels abstract, you can simply look up a skill file online. Once you see one, the idea usually becomes clear very quickly.
总之，可以把一个技能的文件理解成某种意义上“智慧的结晶”。
In short, a skill file can be understood as a kind of distilled operational knowledge.
人们就不需要自己反复的和语言模型去沟通了，只需要告诉他使用这个技能，语言模型就会按部就班的按照这个规范流程去执行了。
This means people no longer need to repeatedly explain the same workflow to the language model. They only need to tell it to use the skill, and the model can then follow the defined procedure step by step.
那，一般谁来写这个SKILL文档呢？
Who usually writes the SKILL document?
可以自己写，也可以让语言模型来写（现在大部分模型都支持自己创建SKILL文档了）。也可以语言模型写完了，然后人再修改也可以。
You can write it yourself, or ask a language model to write it. Most models now support creating SKILL documents. Another common approach is to let the model draft it first and then revise it manually.
使用别人的SKILL
Using someone else’s SKILL
如果你想和其他人交换SKILL，也是可以的。只需要把对方的文档复制过来，放到正确的位置，就可以使用了。
If you want to exchange SKILLs with other people, that is also possible. You simply copy their document into the correct location, and then it becomes available for use.
但需要注意的是，来路不明的SKILL要慎用，因为你不知道它里面可能会安装或者有什么奇奇怪怪的东西，但是又不容易发觉。
However, untrusted SKILLs should be used with caution, because you may not know what they install or what unusual behaviour they contain, and such issues may not be immediately obvious.
前面我们提到过，OpenClaw会构建一个System prompt。如果我们本地有SKILL.md文件的话，那么它就会把里面的必要的内容也加入到这个System prompt里面。
As mentioned earlier, OpenClaw constructs a system prompt. If there are local SKILL.md files, it will include the necessary information from them in that prompt as well.
比如，目前我们有什么技能可以使用，如果需要的话，可以去哪里读取这些技能的具体内容。
For example, it may include which skills are currently available and, if needed, where their full contents can be read.
毕竟，每个SKILL的具体内容可能很长很长，我们也没有必要在不使用他们的时候，就把这些具体内容流程都塞到System prompt里面。这样做的话，可以对资源和自己的金钱极大的浪费呀！
After all, the full contents of each SKILL may be very long, and there is no need to place all of that detail into the system prompt when the skill is not being used. Doing so would be a major waste of both resources and money.
这种先不载入Skill的具体内容，只载入他们的简要信息的按需读取的做法，其实也是一种意义上的Context Engineering。
This approach of loading only brief skill metadata first, and reading the full content only when needed, is also a form of context engineering.
11. AI Agent/OpenClaw如何记忆 (How AI Agents (OpenClaw) Handle Memory)
我们已经知道了，每一次和OpenClaw实际上都是在做token的累积。和他聊的越多，那么它每次和语言模型在沟通时，产生的聊天内容也就越多。
We already know that every interaction with OpenClaw adds to the total token usage. The more you talk to it, the more conversational context is accumulated each time it communicates with the language model.
但，语言模型的Context Window的大小是有限的，我们不可能永远聊下去；另一方面，就算理论上支持很长很长，我们很难触及的Context Window的大小，我们和他说的内容越多，可能与以前已经聊过的内容相关的任务，他也可能会表现的越来越差，等等。
However, the context window of a language model is limited, so a conversation cannot continue growing forever. Even if the theoretical limit is very large, performance on tasks related to earlier parts of the conversation may still gradually decline as more content accumulates.
总之，我们有的时候要开启新的对话。那OpenClaw是怎么记住以前的对话的呢？
So, sooner or later, we need to start a new conversation. The question is, how does OpenClaw remember what happened before?
OpenClaw会自主维护一个自己的记忆档案。简单来说，OpenClaw有写一本属于自己的日记（Daily notes），记录了比较原始的内容，每次对话都发生了什么；也有一个文件用来存储长期记忆（Long-term），可能是一些关键性的事件等等。
OpenClaw maintains its own memory records. Put simply, it keeps a kind of diary through daily notes, which record relatively raw details about what happened in each interaction. It also maintains a long term memory file for more important events and persistent information.
什么时候写进自己的记忆档案？
When does it write something into memory?
用户可以明确的让OpenClaw去记录：比如“把刚才发生的事情你记下来”、“你的生日是X月X日”等。它会自己去找到的对应的文件，然后对其进行更新；
The user can explicitly ask OpenClaw to record something. For example, “Please remember what just happened” or “Your birthday is on this date”. It will then locate the relevant file and update it accordingly.
OpenClaw自己也会决定何时记录：这种“何时”的信息藏在了system prompt里面，以及藏在这个prompt里面的信息是从AGENTS.md中读取的，比如：
OpenClaw can also decide for itself when something should be recorded. The rules for this are embedded in the system prompt, and that information is typically read from AGENTS.md, for the example above.
读取记忆
Reading memory
那什么时候读取记忆，其实也已经在OpenClaw的内部写清楚了：
The conditions for reading memory are also defined internally within OpenClaw.
在回答任何的关于之前已经做过的工作、决定等消息时，要先使用memory_search功能来搜索自己的记录档案，找到与其相关的内容。
When responding to anything related to previous work, earlier decisions, or past interactions, it should first use the memory_search function to search its own records and retrieve relevant information.
这里的memory_search，可以使用RAG（Retrieval-Augmented Generation）来实现。你可能对这个概念已经不陌生了，如果不了解也没有关系，总之就是一个可以根据关键词等，从OpenClaw自己维护的记忆档案中找到相关内容的工具就是了。
Here, memory_search can be implemented using RAG, or Retrieval Augmented Generation. You may already be familiar with this idea, but if not, it is enough to think of it as a tool that retrieves relevant content from OpenClaw’s own memory records based on keywords or semantic similarity.
除此之外，一般OpenClaw在启动时，也会预先自己读取载入长期记忆和最近的日记到system prompt中去。
In addition, when OpenClaw starts up, it will usually preload its long term memory and recent daily notes into the system prompt.
OpenClaw有的时候只是做做样子
Sometimes OpenClaw only pretends to remember
需要注意的是，有的时候OpenClaw跟你说把什么东西记录下来了，但实际上他可能只是敷衍你而已，并没有真正的记录下来（尤其是发生在比较笨的模型下）。
It is important to note that sometimes OpenClaw may tell you it has recorded something, while in reality it may not have done so at all, especially when the underlying model is less capable.
所以，对于很重要的内容，需要你自己检查一下！
If the information is important, you should verify it yourself.
12. AI Agent/OpenClaw如何定时工作 (How AI Agents (OpenClaw) Perform Scheduled Tasks)
心跳（Heartbeat）机制
The heartbeat mechanism
语言模型如果没有什么事情是不会主动和你说话的，但我们可以用心跳机制让其主动和我们对话。
A language model will not usually speak to you on its own if nothing is happening, but a heartbeat mechanism can be used to trigger proactive behaviour.
“心跳”的含义其实就是每隔一段时间，我们就戳一下OpenClaw，让他自主执行一些任务。比如定期检查是否有新邮件等等。其次，这个任务也不一定是非常具体的任务，你也可以下达比较模糊抽象的任务，比如“你可以让自己向着自己的人生目标想办法更前进一些”等。
A heartbeat simply means nudging OpenClaw at regular intervals so that it can carry out tasks on its own, such as periodically checking for new emails. The task does not always need to be highly specific. It can also be abstract, such as asking it to find ways to make progress towards its broader goals or "life"-goals.
Cron Job还有一种心跳系统是Cron Job，其实也是安排定时任务。是属于上面“规律心跳”之外的额外的心跳任务。当到时间时，就会执行，比如“每天中午做一个影片”等。
Another scheduling mechanism is the cron job, which is also used for timed tasks. It can be seen as an additional trigger beyond regular heartbeat checks. When the scheduled time arrives, the task is executed, for example, creating a video every day at noon.
Cron Job还有一个功能就是，它会让OpenClaw学会等待。比如我们想让它去使用某个网站去生成一个影片、或者一个Slides，但极大可能会发生的是，不太可能即时完成。
A cron job also helps OpenClaw learn how to wait. For example, if it uses a website to generate a video or a slide deck, the result is unlikely to be ready immediately.
那我们可以通过让OpenClaw或者在它的memory中写到“当出现需要等待的任务时，你每隔X分钟就去看一下，是不是做好了”具备了等待的意识。然后它就会通过Cron Job系统来让它自己过一段时间后再去检查有没有做完。
In that case, we can instruct OpenClaw, either directly or through its memory, that whenever a task requires waiting, it should check again every X minutes to see whether the result is ready. It can then use the cron job system to schedule those follow up checks.
13. AI Agent/OpenClaw如何长时间自主运行 (How AI Agents (OpenClaw) Run Autonomously Over Time)
前面我们已经提到了，我们和OpenClaw聊的越多，那么这个Context就会越来越多！而当长到一定程度时，我们最好对Context中的一些内容进行压缩。
As mentioned earlier, the more we interact with OpenClaw, the larger the context becomes. Once it grows beyond a certain point, it is beneficial to compress parts of the context.
比如，可以让语言模型帮我们把很多很多轮的信息总结压缩一些，只保留其中关键的信息。如果后续又聊的越来越多，已经被压缩的内容，也是有可能再次、多次被总结压缩的。
For example, we can ask the language model to summarise multiple rounds of conversation and retain only the key information. As the conversation continues to grow, previously compressed content may be summarised again, possibly multiple times.
当然压缩的策略是多种多样的，这并不是唯一的解法。还有一种思路是，一般如果在对话过程中使用了工具，那么大多数时候，我们其实更关心这个工具的产出结果，而不关心中间到底发生了什么。所以，这个思路是，可以对Tool Out下手。
There are many strategies for compression, and this is not the only approach. Another idea is that when tools are used during interaction, we are usually more interested in their outputs rather than the detailed intermediate steps.
比如在OpenClaw中，有两种精简Context的策略，一种是Soft Trim，另一种是Hard Clear。
In OpenClaw, there are two strategies for reducing context: Soft Trim and Hard Clear.
Soft Trim: 只保留开头和结尾，往往这两个位置有最重要的信息。
Soft Trim: only the beginning and the end are kept, as these sections often contain the most important information
Hard Clear: 把整个工具的输出都删掉。
Hard Clear: the entire tool output is removed
当然了，这些策略并没有绝对的哪个好、哪个不好，更重要的是，在你使用的过程中，哪个策略对你的日常使用感受中更吻合。这可能就需要自己去尝试啦。
These strategies are not inherently better or worse than each other. What matters is which one fits your workflow and usage experience best, and this usually requires some experimentation.
收尾 Final Thoughts
理解 AI Agent 的关键，并不在于掌握某一个具体工具，而在于理解它是如何在模型、工具、记忆与上下文之间不断权衡与协作的。
Understanding AI agents is not about mastering a specific tool, but about grasping how they continuously balance and coordinate between the model, tools, memory, and context.
