Architecture
The Colleague Who Forgets Every Morning
Today’s legal AI reprints itself at every interaction. A brilliant junior who forgets each night what he learned is not precious. He is exhausting.
There is, in the way legal AI presents itself today, an obsession with demonstration. Each interaction is treated as a new occasion to impress: a new question, a new answer, a new linguistic feat. The model restarts from zero, mobilizes all its capabilities, produces something plausibly remarkable, then forgets. At the next question, it will begin again, blank, as if nothing had preceded.
This mode is excellent for a demonstration, where each exchange is isolated and the sought effect is the one-off dazzle. It is bad for real legal work, which is not a succession of isolated exchanges but a continuity, the same matter taken up day after day by the same people. What makes a good demonstration and what makes a good work tool are not only different: on this precise point, they are opposed.
What impresses in the demo is that the model starts from zero. What exhausts in use is exactly the same thing.
The fatigue of re-explaining everything each day
Because what a professional expects from a tool used every day, matter after matter, month after month, is not a succession of impressions. It is a stable presence, that remembers, adapts, progresses with them. The repetition of the feat, far from reassuring, ends up wearying, because it comes, each time, with the obligation to lay everything out again, explain everything again, recontextualize everything. The first day’s dazzle becomes, by the hundredth, a chore.
Take the image of a very brilliant junior who, every morning, re-impresses those around him having forgotten everything from the day before. The first day, you are struck by his sharpness. The second, you find it again with pleasure. The tenth, you begin to understand that you will have, every day, to re-explain the matter to him, remind him what was decided, point out again the client’s preferences and the pitfalls already met. His brilliance never accumulates; it replays identically, and that is precisely what makes it exhausting. You do not entrust him with the important matters, not because he is incompetent, but because you cannot lean on him.
A genius who forgets each night is not an asset. It is a burden disguised as talent.
One must dwell on that word, exhausting, for it says something precise. A tool that produces bad answers, you reject quickly and move on: the problem is clear, the decision simple. A tool that produces excellent answers but forgets everything between uses poses a more insidious problem, because it gives, each time, the impression of being on the verge of being precious. You hold on, re-explain, reinvest, and exhaust yourself in this never-capitalized effort. The good amnesiac tool is more tiring than the bad frank one, because it sustains hope while demanding, each time, that you start from zero with it.
The fatigue at issue is not a convenient metaphor; it describes a real and measurable experience. Re-explaining a context costs time and attention, and that cost repeats at every interaction. But there is worse than lost time: there is the distrust that sets in. You do not grant your trust to someone who forgets, because trust supposes you can lean on a continuity. A tool that forgets places the user in the position of never being able to truly rely on it, and this permanent vigilance is, in the long run, more exhausting than the absence of a tool.
The colleague you keep, and the one you avoid
One can extend the image, for it illuminates an often-neglected point about what truly makes a colleague’s value. What makes that value is not that he is smarter than a brilliant beginner; it is that he carries within him the history of the matters, the why of past decisions, the client’s preferences, the pitfalls already met. This accumulation, invisible in a job interview, is precisely what makes it so that after three years, you entrust him with what you would entrust to no one else. It is often worth more than raw sharpness, because it is what cannot be replaced overnight.
Raw sharpness impresses in the interview. It is accumulation that makes the value after three years.
This preference for accumulation is not a firm’s bias; it is the very structure of legal work. A matter is never handled at once: it builds by layers, over weeks or months, and each layer supposes the memory of the previous ones. A decision taken in March commits what you write in September; a position settled for a client orients others. In such work, forgetting is not a minor flaw compensated by brilliance: it is disqualifying, because it breaks the continuity on which everything else depends. This is why a reliable but ordinary colleague is, on a long matter, preferable to a genius who erases everything each night.
An AI that reprints itself bets everything on sharpness and abandons accumulation; it chooses, without saying so, the least useful profile over time. It resembles that colleague you keep for occasional flashes but to whom you entrust nothing ongoing, because you know you will have, each time, to start over. The legal AI market has spent two years recruiting, in sum, the dazzling and amnesiac junior, and being surprised that organizations do not entrust him with their real work.
What a firm seeks, by contrast, is the colleague who stays: the one whose knowledge of the matter deepens, whose answers take account of all that preceded, whom you need not distrust because he erases nothing. Such a colleague does not give demonstrations; he does something far more precious and far less spectacular: he becomes, over time, someone you can count on. It is that quality, reliability over time rather than brilliance in the instant, that distinguishes a tool you really use from a tool you admire then abandon.
It is in this direction that MAX was conceived: not to produce each time the best possible demonstration, but to become, in use, the colleague who remembers. This supposes accepting a trade-off few vendors own: to impress less at first contact in order to serve better at the hundredth. A product judged over time gives, by definition, a less dazzling demonstration than a product judged in the instant; but it is real use, not first impression, that ultimately decides an infrastructure’s value.
The first interaction is not the decisive moment. It is the hundredth, when the demonstration has become a familiarity.
This reversal of criterion, from first contact to the hundredth, is not merely a matter of commercial patience. It says where a work tool’s value really lies. A tool judged on its first demonstration is judged on its most superficial aspect; a tool judged on its hundredth use is judged on what it has become for the one who uses it. And what a tool becomes is precisely what a demonstration cannot show, because it does not yet exist at the moment the demonstration is made.
The consequence is counterintuitive in demonstration and obvious in use: MAX’s value is appreciated over time, not in the instant. An AI that re-impresses every morning stays a product you admire and abandon; an AI that remembers becomes a colleague you lean on. And it is in the second, not the first, that lasting legal work resides.
