Architecture
Sessions Versus Matters: The Architectural Gap
There is a quiet mismatch at the heart of legal AI. Naming it explains almost everything practitioners find frustrating without being able to articulate, and forces a reconsideration of what one should expect from a product.
Most of the frustration lawyers feel when they use AI tools is real, specific, and oddly difficult to articulate. The tool is clearly capable. It drafts well, it answers well, it impresses in the demonstration. And yet, in daily use, something is persistently wrong. The tool forgets. It contradicts itself across a week. It treats every interaction as if it were the first. The lawyer ends up doing the very work of continuity that the tool was supposed to take away. There is a name for what is happening here, and once you see it, you cannot unsee it.
As long as a phenomenon has no name, it remains a diffuse discomfort, attributed sometimes to the inexperience of the user, sometimes to the immaturity of the technology. Once named, it becomes a category, and that category makes it possible to reason clearly about what must be built to solve it. The distinction that follows, between sessions and matters, is that name.
Two primitives, two ways of inhabiting time
AI models operate in sessions. A session is a bounded conversation: it opens, you work inside it, it closes, and when it closes the context evaporates. This is not a flaw in the models. It is simply what they are. A session is the natural unit of a language model, the frame within which it reasons. Outside that frame, by default, there is nothing.
Legal organisations do not operate in sessions. They operate in matters. A matter is the opposite of a session in almost every respect: it is long-lived rather than bounded, layered rather than flat, and it carries decisions, versions, positions, history and accountability across months and across many people. A matter has a memory by necessity, because the work depends on what was decided three weeks ago by someone who has since gone on holiday. A matter is not a conversation. It is an institution in miniature.
Models think in sessions. Law lives in matters. Almost every frustration with legal AI lives in the distance between the two.
This opposition of primitives has a philosophical consequence before it has a technical one. The session inhabits the present. It begins, it unfolds, it ends. It is compact and self-contained. The matter inhabits duration. It begins without anyone knowing how long it will last, it thickens progressively, it goes through changes of team, of strategy, of legal context, and it concludes, sometimes, several years after its opening. Asking an object that inhabits the present to hold an object that inhabits duration is asking a snapshot to replace a film.
Where the seams show
When a session-based tool is dropped into a matter-based world, the seams show immediately, and they show in ways that are quietly corrosive rather than dramatically broken. The assistant forgets what was decided last week, so the lawyer re-explains. It cannot see why a clause was drafted the way it was, so it cheerfully suggests an edit that reopens a point that was settled, painfully, a month ago. It treats the fortieth interaction on a matter with exactly the same blank-slate freshness as the first, which feels less like working with a colleague and more like training a brilliant amnesiac over and over again.
A tool that forgets does not remove the work of remembering. It hands it back to the lawyer, disguised as assistance.
This primitive error has very concrete manifestations every practitioner has already encountered. A lawyer picks up a matter after three weeks and has to re-explain everything to the tool, because the session from three weeks ago no longer exists. Two members of the same team work on the same matter with the same tool, but each in their own session, and the tool is unaware that one contradicts what the other established yesterday. An archived matter is reopened a year later, and no usable trace of the AI assistance remains. Each of these situations is a symptom of the same primitive mismatch.
The most deceptive thing is that this mismatch is invisible precisely where tools are evaluated. A sales demonstration fits inside one session: you ask a question, you get a brilliant answer, you are convinced. The matter only reveals itself over time, when it thickens, changes hands, crosses weeks. This is why so many tools impress in the demo and disappoint in use: the demo tests the primitive that suits them, real use tests the one they lack.
What the mismatch costs economically
The mismatch between sessions and matters has a measurable economic cost that appears in no budget because it lodges in the interstices of work. When a lawyer has to re-explain the context of a matter to a tool because the previous session is lost, that time consumes billable time. When two associates do not know that their same tool has given them different answers on the same matter, the gap is caught by an additional layer of human review. When a matter reopened a year later carries no trace of previous AI use, the reconstruction work that follows is not charged to the tool.
This cost is precisely the kind of cost that does not trigger alarms in an organisation. It breaks nothing visibly. It slowly dilutes the productivity that technology was supposed to bring, until a point where one observes that the hours gained by AI on generation are exactly offset by the hours lost on continuity. At that moment, the return on investment of legal AI seems strangely zero, without one being able to say why. The reason is that one has paid twice: once for generation, and once for the continuity that generation did not carry.
What the mismatch costs legally
Beyond the economic cost, there is a question of responsibility the profession will not be able to dodge for long. The coherence of a matter is not just a quality of service; it is a professional expectation. A lawyer who takes contradictory positions on the same point within the same matter engages their credibility, and possibly their responsibility, independently of the pointwise quality of each position taken. If the tool that assists this lawyer is, by construction, incapable of remembering the position taken three weeks ago, the risk of contradiction is not an accident, it is a structural property of the use.
A firm that broadly deploys tools based on the session primitive, without a layer above to hold the matter, tacitly accepts to migrate the responsibility for the coherence onto its individual users. As long as they are vigilant, all is well. When a user does not have the historical context of the matter, or when a matter changes hands, coherence can fall, and no one will know precisely who was supposed to carry it. This diffusion of responsibility is the kind of arrangement that ends, sooner or later, in producing a serious incident.
Why this is an architecture, not a feature
It is tempting to think this gap could be closed with a feature. Add a memory function. Bolt on a history tab. Let the user paste in context at the start of each session. But these are patches over a structural mismatch, and they fail in the same way for the same reason: they ask the human to do the work of bridging two incompatible models of time. The real answer is not a feature added to a session-based tool. It is a layer that thinks in matters from the ground up.
Such a layer holds the matter as the primary unit. It remembers across time without being asked. It knows why, not just what. It governs who did what, and it preserves the line of reasoning that makes a matter coherent rather than a pile of disconnected interactions. The models still do the generating; they are good at that. But the layer above them does the holding, and the holding is what turns scattered AI use into legal work that stands up.
Closing the gap between sessions and matters is not a feature. It is the product.
This is the difference between a tool that is impressive in a demonstration and a system that is trusted in a practice. The demonstration lives inside a single session, where the mismatch never has time to appear. The practice lives in matters, where the mismatch is everything. Build for the session and you win the demo. Build for the matter and you win the work.
