
Imagine you have hired an exceptionally smart employee, someone who can do almost anything you ask. There is only one problem: he has no long-term memory. He is a bit like Dory in Finding Nemo, only in reverse: Dory forgets what just happened, while your genius employee forgets what happened before — your last conversation, last week’s meeting, a promise a client made a month ago — or at least he cannot remember it very well. Hand him a task cold and he might get it exactly right, or he might completely mess it up. But he is an astonishingly fast reader, and he holds on to what he has just read remarkably well. Give him the documents on your engineering projects and he can take on a surprising amount of engineering work; give him the marketing documents and he can do the same for marketing. This is the way I have found most useful to think about AI.
Some of my best meeting preparation now happens in the car. On the way to a client, I talk the meeting through with Claude, out loud, the way I might once have done with a colleague in the passenger seat. Claude knows what I am talking about, because it is connected to my calendar, email and my company information. It has the background on the client, our previous conversations and the goal of the meeting, and it knows which information about that client is sensitive and is meant to stay inside that client’s own workspace.
On one of those drives, it told me to remind the client about an introduction they had promised me about a month earlier. I would almost certainly have forgotten. And an introduction a client has already offered you is about as warm as a lead ever gets.
Should we say that a smart AI agent remembered an important thing that I might have missed otherwise? But that is not quite what happened, and the difference is the whole point. Before our conversation started, a small script had put together a session brief, a short document that included everything we had agreed to with that client in the past. Claude read it. The model did not remember the promise. The system did.
It is worth being precise about what kind of memory you are being offered, because there are really three kinds, and they overlap more than you might think.
The first is personal memory. Claude, ChatGPT and other assistants now remember things about you across conversations and can look back through your past chats: your role, your preferences, what you were working on last Tuesday. It is helpful, and it is a memory of your conversations with one assistant, not a memory of your company.
The second is project memory. The leading assistants let you set up projects: documents, instructions and conversations kept together, sometimes connected to your drives, and increasingly with a memory of their own. That is a real step up; it is the equivalent of handing your genius employee the right binder every morning. What a project does not give you on its own is the rest: a record of where each piece of knowledge came from, rules about who may change it, walls between clients, the freedom to switch assistants without losing it, and a deliberate way to write back what each job taught you.
The third is a full-blown harness, and it is not another memory feature at all; it is the system around the model. Many of you are familiar with AI harnesses, but many are not, and for those who are not, it helps to think of the harnesses you already know. A climbing harness does not do the climbing; it ties the climber to the rope, the anchors and the person holding the other end, so that a strong climber can attempt something hard without one slip turning into a fall.
An AI harness connects the model to what your company knows, and helps it decide what it reads before it starts work, and what happens to what it learned once the work is done. It lives in your company’s own systems rather than inside any one assistant’s account. It takes in new material as it arrives — meeting notes, emails, decisions, client feedback, any market evidence — puts together a fresh brief before each job, is meant to keep each client’s information behind its own wall, and writes what was learned back when the job is done. Getting the right material in front of the agent before it starts is the part everybody notices. Writing what it learned back into the company afterwards is the part that makes the business work. Inside a single project, your genius employee no longer starts from scratch every morning; with a harness, the whole organization can get smarter with every piece of work, not just that one project. A project remembers the work. A harness governs it.
Here is what I find most interesting about it. As the frontier models like Claude and ChatGPT get better, in my experience one part of the harness keeps getting thinner: the specific instructions, the workarounds, the careful hand-holding the agent used to need. Not long ago, getting a good result meant spelling out every step, pasting in examples of a good answer, reminding the agent to check its arithmetic and not to make things up, and chopping a long job into small pieces so it would not lose the thread halfway through. More and more of that, the models now handle on their own.
As the generic instructions of the harness keep getting thinner, there is another part that quietly keeps growing — those are your evidence, your decisions, your outcomes and the measurements of how your company performs. It works like this. Every piece of work leaves a trace: what the client asked for, what you delivered, what it cost, how long it took, what the client said afterwards and what happened next. One trace tells you very little. Hundreds of them, kept in one place and read by something that never gets tired, start to surface patterns you might never have noticed yourself. Some will be coincidence; the ones worth acting on are the ones you check. As you learn from your own workflows, your knowledge of your market, your customers and your own company keeps growing.
That is where the differentiation lives. Anybody can run any agent; the tools are available to everybody. The difference is whether all that output makes your firm smarter every single time: whether someone looks at what the work taught you, decides which lessons should travel, and changes how you work because of them. That is how you leapfrog the competition.
A good example of this is older than any of the tools. Zara built much of its business on a simple loop: store managers listen to what customers ask for and cannot find on the racks, and pass it back to headquarters in Spain, where the designers act on it. In one widely reported case from 2016, store managers relayed that customers wanted more hardware on their clothes, and a coat designed on that feedback went from idea to store in about 25 days, at a time when the traditional apparel pipeline took months. The feedback alone was not the secret; Zara had built its design, production and logistics to act on it quickly. What it had was a faster loop between what its customers said and what it did next. I think an AI harness is making the learning half of that loop affordable for a company of almost any size.
So the big question for any business is a simple one: where does your knowledge live — in chats, or in your company brain?
We call our firm’s harness the Knowledge Engine, and if building it has taught us one thing, it is that the major part is capture: not only meeting notes, emails and messages, but pretty much everything that happens in the company — decisions, market evidence, events, verbal agreements and so on.
What does all of this buy us? Speed, first of all: I can prepare for a meeting in the time it takes to drive to it, because nobody has to brief the agent from scratch. Efficiency, because the agent arrives already knowing the client. And capacity: a small team can keep track of far more clients, commitments and projects than it ever could before, and get through the work a lot faster.
The same harness that remembers can also learn, and feedback is one of the most valuable things you can feed it. Clients tell you what landed and what did not, which demos were awesome and which never struck a chord. None of it is universal; what one client loves, another shrugs at. But when all of it goes into the harness, each new deliverable starts from what the previous ones taught you, and in my experience that shortens the learning curve dramatically. Sometimes a single comment reaches well beyond the client who made it. After one of our presentations, a client asked to see exactly how we run a particular workflow ourselves: answering customer requests, with an agent drafting each reply and a person approving it before it goes out. That demo now opens our presentations.
So if you want this kind of AI employee, where do you start? Not with the model.
The very first step is to bring everything related to a client together: SharePoint, Teams, email, spreadsheets, Google Drive, Dropbox, Box and, hardest of all, people’s heads. Sometimes that means moving it into one place; when a client needs its material to stay in its own systems, it means connecting to it where it lives, with the right permissions. Either way, the agent gets one coherent place to start from, which is much easier, and in my experience cheaper, than having it search half a dozen applications every time it starts work, because all that searching uses up tokens. And if you serve multiple clients, it means one such place for each of them, with a wall in between.
Then start small, and think of it as a trust ladder. You build trust with an agent the same way you build it with a new employee. At first it only suggests; then it drafts; then it acts inside the company; then it acts outside the company, with your approval; and only at the top does it act on its own and report back. My rule of thumb: when it handles a relatively simple task consistently three times in a row, that is my cue to consider trusting it with a little more of that kind of work — with your sign-off, not its own. Moving it up to a different kind of authority, such as acting outside the company, takes more evidence than three good results. When it fails, it steps back down. Three is a rule of thumb for everyday work, not a safety standard for work where a mistake could hurt someone. A sensible place to draw a line, at least for now, is before an agent sends automated emails or messages to clients without your approval; that is where we currently draw ours.
And keep the judgment for yourself. The harness reminded me about the introduction; deciding whether to raise it, and how, was still my job.
If all of this sounds vaguely familiar, you may have seen 50 First Dates. Drew Barrymore plays a woman who loses each day’s memories overnight, and Adam Sandler, rather than give up on her, makes her a videotape to play the moment she wakes up, so she can catch up on her own life. It is the most romantic session brief ever recorded. It also got him the girl, which is more than I can promise for your CRM.


