Every organization has two kinds of help.
There is the bright new hire with a perfect notebook. Ask her a question and she opens it, finds the right page, reads what was decided last spring, and gives you a good answer.
And there is the colleague who has been there ten years. Ask her the same question and she doesn’t look anything up. She already weighs it the way you would. She knows which client you’d turn down, which shortcut you’d never take, which detail matters to you and which is noise.
Both are useful. Only one of them gets better at being you.
From prompt machine to agent with a home
I have worked with AI agents every day for months, and I have watched them grow up fast.
OpenClaw was the first one that felt personal. An open-source agent on your own machine, reachable from WhatsApp or Telegram, with scheduled tasks, tools and a memory that survives the conversation. I wrote about it in February: for the first time, an AI already knew who I was. Hermes Agent from Nous Research went a step further, with a built-in learning loop. It writes down what it learns about you and turns things that work into reusable skills. And with Grok Bot the agent got a home of its own: a persistent computer in the cloud, a team of bots that keep working after you close the laptop, and the cloud agent as the normal way work gets done.
Look closely and all three solve memory the same way. With notes. Markdown files, memory files, skill files. The model reads them at the start of the job. In OpenClaw’s own documentation, the model “only remembers what gets saved to disk.”
That is the new hire with a better notebook. And I keep hitting the same wall. More context works, until it doesn’t. The notebook gets thicker, and somewhere along the way the agent gets slower, vaguer, more generic. It reads more and understands less.
A note records what happened. Only a trained model changes how the next decision is weighed.
What only a human brings
Elon Musk has named three things he considers most important for AI: truth, curiosity and beauty. I read them as the three things humans bring that machines can’t produce alone. Truth is contact with reality, the sensor that catches an error before it compounds. Curiosity is the choice of where to look next. Beauty is taste: the judgment about what is worth making, keeping or doing first.
That is where I want my attention to go. Yet even good agents still treat me as the technical operator of last resort. So I tell them: handle everything yourself, safely and without breaking things, and come to me only for what only a human can give. What is actually true out there. Where to look next. What is worth it.
A general model can follow that instruction for a while. It stays followed only when the model knows me, and keeps learning.
Why a thicker notebook isn’t enough
The standard answer to memory is RAG, retrieval-augmented generation: search your documents at the moment of the question and paste the best passages in front of the model. It works, and it is everywhere. But every answer starts from scratch, and the model only ever sees fragments.
In April, Andrej Karpathy described something better: the LLM wiki. Raw sources go in, untouched. The model compiles them into a living, interlinked wiki, with pages for people, ideas and decisions, and good answers get filed back. The knowledge is compiled once and kept current, instead of being dug up again on every question. I run one. It is the best notebook I have ever had.
But Karpathy himself pointed to the next step: generate synthetic training data from that wiki and fine-tune, so the model knows the material in its weights instead of only in its context window.
That is the step that matters. Every RAG system and every wiki still has to select the right pages and slide them in front of a generalist. Cost grows, speed drops, the signal gets diluted. The model never becomes the local expert. It only gets better at reading about one.
Turn the notebook into the colleague.
The view from the lab
Thinking Machines Lab made the underlying case in July, in The Future Worth Building Is Human. Most AI today is trained in a handful of places and then frozen. But the knowledge that runs a business is tacit and local, and people doing the work create more of it every day. Chess and mathematics are the exceptions: the goal is fixed and the board is visible to everyone. Your business is not chess.
Their answer is Tinker, a service for post-training. You bring the data, the training loop and your definition of success. They run the machines. The technique is LoRA, low-rank adaptation: instead of retraining a whole open model, you train a thin adapter on top of it. That adapter holds your expertise, and you can download it. It is yours.
The results so far are striking. Bridgewater had expert investors label financial documents and trained an open model on their judgment. It got roughly 85 out of 100 right where the best frontier model they tested got 78. That is about a third fewer mistakes, at around a fourteenth of the cost to run. A team at the University of Illinois trained a model to turn questions into database queries that matched human accuracy on a cleaned benchmark, at a fraction of the price of the leading general systems. Exa took a small 4-billion-parameter model and trained it for multi-step search until it beat an untrained model sixty times its size.
The pattern is the point. When expertise goes into the training rather than into the prompt, a smaller model beats a bigger one at the work that matters to the people who own it.
Three layers, all of them needed
Building the colleague takes three things. They are not alternatives, and this is not a hybrid. Each one fails without the other two.
The first is a knowledge base, in the form of an LLM wiki or a knowledge graph: facts, entities, relations and decisions, kept up by agents and confirmed by a human before anything counts as settled.
The second is a small language model of your own: an open model with a LoRA adapter, fine-tuned on your best data, your preferences and your standards. This is where how you weigh things lives.
The third is a synthesis loop with continual learning. Agents pull what is relevant from the knowledge base, your model turns it into something new in your own framing, a human reviews it, and the approved results become fresh training data, synthetic examples and preference pairs. Every so often they get folded back into the adapter, carefully, so the model learns the new without forgetting the old.
Leave out the model and you are back to the new hire with the notebook. Leave out the knowledge base and you get instinct without memory. Leave out the loop and you get the colleague who stopped learning the year she was hired.
The big frontier model doesn’t go away. It stays for the heavy, open-ended reasoning. The agent environment orchestrates the flow. Your own model does what the frontier can’t: answer quickly and consistently, the way you would.
Building it
I have started building this for myself, and it is the most exciting thing I am working on.
The knowledge base is in daily use. The first small training runs on my own preferences are done. One lesson surprised me. Asking which of two things I preferred taught the model far more than asking me to score things one by one. Pairwise preference beats ratings, as the research says, but it is something else to see it in your own data. Another lesson was humbling. A small personal dataset teaches broad instincts first, and it takes more good examples before the model clearly beats a general one at being me. The loop is designed and ready to run. Agents already keep the wiki, draft the training examples and manage the review, and they bring me only the decisions that need a human.
The same pattern fits a business team. Each team gets its own knowledge base and its own adapter, trained on how that team decides. The agents run the loop. The people stay the source of truth, curiosity and beauty.
None of this needs a research lab anymore. It needs a clean wiki, good examples and patience.
What could prove me wrong
Frontier models may learn to personalize in ways I don’t see coming. Privacy, cost and catastrophic forgetting, where a model loses old skills while learning new ones, are real problems. And some will say this is just RAG with extra steps.
So here is the test. Give a large frontier model RAG and a long context window. Give a small model the full stack. Then measure which one makes my decisions the way I would. If the frontier model wins, I’m wrong, and I would rather know.
The colleague
The new hire with the notebook will keep getting smarter, and that’s a gift. But the value in your company, and in your life, was never in the notebook. It lives in the way you weigh things, and nobody else can train that for you.
The shape of the agent is settled now. What’s left is making it yours.