The memory a personal agent needs is the memory a business agent can't have
· Prabhu Eshwarla
Two AI systems can share almost every component and still be built on opposite principles. The clearest example I know is memory.
Nous Research's Hermes is an open-source personal agent. Its headline promise is "the agent that grows with you." It remembers your projects, auto-generates skills from what you do, and never forgets how it solved a problem. That is a genuinely good design for a personal agent, and Hermes executes it well.
Forge, the platform I build, deliberately does not do this. Not because we couldn't, but because for the thing Forge is for, accumulating memory is the wrong choice. Working through why turned out to be a useful lesson in what I'd call harness engineering: the design of everything around the model that decides how it behaves. The model is often the same. The harness is where the product actually lives.
Two kinds of memory
When people say an AI "has memory," they usually mean one of three different things, and it helps to separate them.
There is reference knowledge: the documents and facts the system answers from. There is working context: the materials in play for the task in front of it right now. And there is longitudinal memory: what the system retains about you across sessions, on its own, over time.
The first two are uncontroversial. Nearly every serious system has them. The third is where personal agents and business agents part ways.
Longitudinal memory is what "grows with you and never forgets" describes. The agent watches what you do, keeps what seems useful, and carries it forward without being asked. Over months it accumulates a picture of you that no one explicitly authored. For a personal companion, that is the entire value. The whole point is that it learns you so you stop repeating yourself.
Why it's right for a companion
For a personal agent, accumulating memory is aligned with the user's interest in a simple way: it is your memory, about you, for you. If it quietly remembers that you prefer terse answers, or that a project stalled last spring, the only person affected is you, and you are the one who benefits. There is no third party. There is no audit. There is no one you have to answer to for what the agent chose to retain. The accumulation is the feature, and the informality is fine because the blast radius is one person.
Hermes is built for exactly this, and its memory design follows from it. That is good harness engineering: the memory model fits the user.
Why it's wrong for a business service
Now change one thing. The agent is no longer working for a person. It is working inside a business, on that business's operations, touching that business's customers and records. Suddenly every property that made accumulating memory attractive becomes a liability.
Consider what "the agent quietly remembered something useful" means in a business setting. The agent retained information about a client, or a deal, or an employee, that no one explicitly filed and no one can fully see. Ask three hard questions of that. Who authorized keeping it? Can you show, on request, everything the system holds about a given customer? If the agent acts on something it "remembered," can you reconstruct where that came from? For a personal companion, these questions don't arise. For a business handling other people's data, they are the whole game, and increasingly the law.
A governed business service lives or dies on being able to answer them. The promise is not "the AI is clever." The promise is "we can show you exactly what it knew, what it did, and where a human decided." Silent accumulating memory breaks that promise by construction. It creates state that is unauthored, unaudited, and hard to expose. The very informality that makes it delightful in a companion makes it indefensible in a system a compliance officer might one day inspect.
So the feature that is the point of a personal agent is a defect in a business one. Same capability, opposite verdict, entirely because the user changed.
What a business agent uses instead
Removing accumulating memory does not leave a business agent memoryless. It leaves it with a different, better-governed shape of memory.
Instead of memory the system accumulates, you use knowledge someone curates. A person deliberately adds a document to a knowledge base. That act is explicit, attributable, and auditable. The system is extremely knowledgeable, but everything it knows got there on purpose, by someone accountable, and can be listed on demand. You trade the magic of "it just remembered" for the property that matters more here: nothing is in the system that someone didn't put there and can't see.
You keep working context for the current task, because that is bounded and transient. You keep conversation history within a session, because that is visible and short-lived. What you drop is the one layer that quietly persists across time without an author. In a governed setting, that layer is not an asset you are missing. It is a risk you chose not to take on.
The harness is the product
This is why I keep coming back to harness engineering as the real discipline. The model Hermes uses and the model Forge uses could be identical. The difference in behavior, and in who each can safely serve, comes almost entirely from the harness around it: what it is allowed to remember, what it must show, where a human sits in the loop, what never leaves the tenant.
Hermes made the right call for a companion: accumulate, personalize, grow with the user. Forge made the opposite call for a business service: curate, attribute, expose everything, accumulate nothing silently. Neither is more advanced than the other. They are answers to different questions.
The mistake would be to assume the impressive consumer behavior is simply the target that business software hasn't reached yet. It isn't. "Grows with you and never forgets" is not a milestone on the way to enterprise AI. It is a design that enterprise AI has specific, defensible reasons to refuse.
When you are deciding how much your own AI should remember, the first question is not technical. It is: who does the memory belong to, and who will you have to answer to for it? Answer that, and the harness mostly designs itself.
