♪ Take notes ♪The rules of honesty at this station.
I built my own Jarvis, not because I lacked a smarter chat robot, but because, with AI getting stronger, I needed a system to keep continuity for me: remember who I am, know what I’m doing, and help me manage a growing number of AIs.
Today’s Agent is already strong. It can write codes, check information, make web pages, and do what I used to do for half a day. But when the window closes, it still doesn’t know me; when the project changes, it still asks me to re-circulate the background; and when there are more tasks, I’m stuck in the window, notification and waiting time.
So Jarvis has never evolved to “add another chat box.” It went through three evolutions: remember me, understand me, manage AI for me.
No, Jarvis was: Agent was strong, but he didn’t know me.
I also use Codex, Claude Code and other AI tools for multiple projects. And every one of these Agents can do it alone, and the problem is they don’t know what just happened to the other window or why I’ve made a decision in the past.
So I do the same thing over and over again: introduce myself, explain the project, post the conclusions, and say what is not moving. I just fed one window, and the other window started over. Agent’s on, I’m waiting; while I’m waiting, I open another mission. Soon, the attention was cut to pieces, and I looked like I was pushing a lot of things at the same time, and actually, the mind was full of unturned tabs.
And there’s a more subtle question: AI produces a lot of things, but it doesn’t really sink. A good judgment in a conversation, not to be found a few days later; the pits of a project and the next; I know how much I have done, but it is not clear how these things are linked to a main line.
It is not that the model is not smart, but rather that the system is not continuous.
First evolution: Remember me.
Jarvis’ first thought was simple: since Agent didn’t know me, show it the data I left.
My best personal data is GitHub. Dozens of projects have documented the direction of my technology and interests, and the presentation of records reveals real working habits: I’m a steady output or a short-term explosion, and I like to set the frame first or to change it, and what’s just up and what’s really going on?
Another important source is the diary. I have been writing daily summaries intermittently since two years ago, and when there is a major turning point in my life, I will do the same. The diary is not a résumé or a self-presentation written for others. People do not have to pack themselves correctly in their journals, so it is firsthand. Project files tell AI what I did and the diary is more likely to tell it why I did it.
And then I’m talking to AI. My dialogue with the bean bag has more than 4 million words, which record my long-standing concerns, my absorbed content and my recurring confusion. Besides, there’s local data on computers. When you connect these things, Jarvis at least doesn’t think of me as a new user.
But there’s got to be cold water:**Collecting information does not mean understanding me.**Four million words thrown into a library is only an indication of the size of the library and not of the correct judgement. The information may be outdated and contradictory at different stages. Not to mention that the old phrase in the search is still on my behalf today.
And privacy boundaries cannot be lifted because they know me better. My Twitter chat records are not imported into Jarvis; private data are given local priority; a scan or a conversation does not automatically contaminate long-term memory. AI wants to remember anything, it has to cross the border and confirm. Otherwise, the so-called “second brain” can easily be transformed into an uncontrolled copier.
Second Evolution: Understanding Me
The first edition of Jarvis is more like a personal knowledge base: information comes in, questions come in, answers from information. This stage is useful, but not sufficient. What really affects my daily work is not just “where something exists,” but what follows:
- What matters most now is what principles cannot be sacrificed for speed;
- What projects do I have? Where are the cards?
- What key decisions have I made and why have I chosen to do so?
- Relations and commitments need to be respected, but should not be exposed;
- What is my energy, attention and priorities today?
So Jarvis started moving from a database to a structured memory. The original or original record, the stabilization judgement is a steady judgement, the project status is the project state and cannot all be thrown into the same drawer. It works.morningHelp me compress today into some really important directions.eveningRewind what moves the goal; interim ideas are passed.captureStay and accumulate to a certain extent.distill;requires into long-term memory firstproposalI’ll confirm.apply。
The most important here is not a few orders, but the “proposal” and “writing” have been deliberately removed. Jarvis said, “I think this is worth remembering for a long time.” But it can’t just rephrase who I am because it’s like that. Once the long-term memory is contaminated by erroneous conclusions, the recommendations, summaries and plans that follow will continue in the wrong direction.
At this point, Jarvis is no longer just an RAG question and answer tool. It is beginning to be my focus and decision-making support: a reminder of the current thread, the reasons for bringing back the past and comparing today ’ s changes with long-term goals. It does not judge for me, but it does not have to start from scratch.
Third evolution: manage AI for me
Then, when I really used the Agent development project, the biggest bottleneck changed: not AI didn’t understand me well, but I was becoming an AI project manager.
I have multiple windows open, one at the front end, one at the beginning, one at the running test, one at the waiting deployment. Each window asks questions, and each window requires acceptance. Agent saved me implementation time, but I used it to cut windows, track progress, copy the context.
So I made a very counterintuitive decision:I only open the main window. The rest of the window, let Jarvis go manage.
Now I give the target to the master. The Master should not personally insert all the details, but find the corresponding project manager. The project manager retains the long-term context of the project and then assigns specific tasks to the engineering, audit or publishing roles. A typical link is:
Master Project Leader → Project Execution/ Independent Audit / Issuance and Acceptance
It’s not for a bunch of Agents to have a party in a group chat. It’s more like GitHub’s step-by-step collaboration: the master handout package, the manager’s delivery code or document, the complete evidence in the report, and only a very brief reply. Those in need of review read the product directly and do not move the entire process of reflection over and over again to all threads. This would save token and avoid contamination of the context by unrelated information.
This collaboration must be limited. Makers and examiners try to be as separate as possible, i.e. maker/checker; access is restricted without defaulting that any thread can be accessed; changes to the code are tested and posted on a real line page; there is a problem to know what changes can be made and can you roll back. The more the number of AIs, the less they can rely on “I feel it should be right.”
Jarvis was involved in an ordinary job. Day
In the morning, I asked Jarvis to give the three most important priorities today, based on the latest project status and diary. I will not take it all on my own, but rather delete those arrangements that appear to be reasonable and, indeed, not in my current state and finally confirm the main line of today.
And I’m only going to say the target in the master window, for example, “Add an article to the public Wiki, keep my mouth shut, do a private scan, and check the end of the phone after it’s published.” The director gave it to the director of the project in Wiki. The person in charge first identifies the warehouse and boundaries, then reads the existing information structure, revises articles, constructs and runs the browser tests. When you need to publish it, it submits the code, waits for Pages, and clicks on the real page online.
I don’t need to keep an eye on the progress strip for a while. I can continue to write the next paragraph of the same article, or deal with the judgment that only I can make on the main line today. Attention, work on the same main line, not because “AI is running” opens 10 new pits for itself.
When the person in charge is finished, the dozens of screens will not be put back into control. It only rewards conclusions, submissions, deployments, risks and reporting positions. Control decides whether an independent audit is required on the basis of a return request; the audit finds problems and returns them for repair; and the evidence is sufficient for the results to be submitted to me for acceptance.
I’ll use it again.eveningLooking back at what really moved forward today, where it was busy and where it was worth leaving behind. The new long-term judgement would not be written in secret, but would first be a proposal. The next morning, these things that I identified could become input for the next round of plans.
This is what I want for “automation”: not for Jarvis, who runs my life while I sleep, but for the re-routing, tracking and sorting that is captured by the system, and I’m still in charge of targets, boundaries and final acceptance.
It’s evolved by failure.
Jarvis wasn’t designed at the table once. Many rules come from real failures.
Title and PermissionIt is because the threads are misidentified and can also be taken over with the wrong permission. Agent’s confidence doesn’t mean it’s in the right project. First, it’s much cheaper to say who you are, what you can do, then to do it.
Reports and replies receivedIt is because “document completion” is not the same as “mission completion”. The officer-in-charge simply leaves the report in the corner, without informing the Control, who does not know whether to proceed with the inspection or continue to wait. Now the whole process is reported, with a very brief conclusion that the absence of either is not closed.
Protection of dirty work treesIt’s because real projects are often modified. Agent, if it were to use the workspace as clean paper, was likely to be able to do what his mission covered. To look first at the state, to submit only the documents for which they are responsible, is the bottom line of collaboration, not the brevity.
Vault and release scansIt is because personal systems are naturally close to the most private data, and open warehouses are naturally duplicated. Core competencies, operating records and private memory must be layered; every time they are packaged or published, they scan the machine ’ s path, account number, key and content that should not be disclosed. “I did not knowingly divulge” is far from sufficient.
These rules look a little stupid, but reliable systems are often a bunch of stupid rules that grow out of failure.
It can’t do anything.
Jarvis won’t decide my life for me, and he won’t be given the ultimate power to explain myself by reading a lot of information.
Its understanding may be incorrect, its long-term memory may be outdated and its structure may be followed by the loss of the original story details. The codes, articles and reports presented by Agent still need to be checked and accepted; multiple roles do not automatically amount to multiple genuinely independent views, and they may be led by the same wrong premise together.
More realistically, the complexity of the system itself becomes a burden. Threads, door closures, reporting, memory levels need to be maintained. If I spend more attention than it saved me to manage Jarvis, it’s a failure. Tools cannot be granted permanent rights because they are well structured.
Final judgment: I don’t want to know my chat machine better. People
I used to say that after feeding GitHub, Diaries, and more than four million words in AI conversations, no one in the world knows me better than Jarvis. Now I think that’s only half right.
“Know me” is not the end. A model can understand me very well today, and tomorrow the context is changed and continuity is broken again. What really matters is whether I have a mechanism to preserve the original facts, to distinguish between ad hoc ideas and long-term judgment, to continue the context of the project, and to limit how I work for the ever-increasing AI.
So Jarvis wasn’t “the best talking robot I know.” It’s a layer of continuity, attention and collaborative control that I built myself in the AI era.
It’s not about turning me into a more busy AI administrator, but about returning my attention to things that can’t be outsourced: judgment, writing, relationships and life.
See the storage design of the system.My, AI, memory system.I don’t know. It brings about a cognitive re-examination.Three-layer capability model。