Money has accounting. Operations have ERP. Relationships have CRM. People have identity. Each acquired a record the moment it began to matter. Intelligence now matters, and it is running without one.
Current AI systems can generate. They can reason. They can act. What they cannot do is hold a position on the world, reliably, verifiably, over time. Five things slip through every session boundary:
Facts about the world, held as claims, not as text fragments scored by similarity.
The interval over which a fact held in reality. Most systems store only “now.”
Knowledge arrives late. A record that cannot say when it learned rewrites its own past.
Who was entitled to see, decide, and act, at the moment it mattered, not as configured today.
The consequence of every action, folded back into the record. Without it, nothing compounds.
Each session reconstructs the world from fragments. Yesterday’s correction is overwritten by last month’s assumption. A model confidently cites a state of affairs that ended in March. Nobody can say what the system believed on the day a decision was made, so nobody can audit the decision.
The failures are not exotic. They are bookkeeping failures, the kind every other consequential domain solved by building a system of record.
With time and boundaries, continuity becomes reliable, decisions become safer, and intelligence compounds.