Shortly after 2:14 on the afternoon of August 14, 2003, the alarm system in FirstEnergy’s grid control room in Ohio stopped working. No more alarms reached the consoles. The operators, investigators later found, did not know they were working without them.
At 2:32 a neighboring utility called about a line that had tripped. FirstEnergy had no alarm and no log of it. The operator, according to the U.S.-Canada task force that investigated, dismissed the information as either not accurate or not relevant to his system.
Elsewhere in the company, at 2:41, the server hosting the alarm function failed, and an automatic page went to FirstEnergy’s IT staff. At 2:54 the backup failed too, and another page went out. The IT staff, the task force found, did not notify the control room. At 3:08 they restarted the main server, believed it was fixed, and didn’t check with the operators. The alarms were still dead.
The IT staff knew the servers had gone down. The operators had the phone calls, and a silence they didn’t know was a fault. Until 3:42, neither group, as far as investigators could find, told the other what it knew. Then a control room operator told IT about the alarm problem.
Candles in a blackout
The psychologist Daniel Wegner opened a 1991 paper with a line from Samuel Johnson: “Knowledge is of two kinds: we know a subject ourselves, or we know where we can find information upon it.” His own example: one partner in a couple may not know where the candles are, but can still find them in a blackout by asking the other.
Wegner and his colleagues had introduced the idea in 1985. In the 1991 paper’s words, transactive memory is “a shared system for encoding, storing, and retrieving information.” Each person takes responsibility for some things and keeps track of who holds the rest.
A team can remember more than any one member, as long as everyone knows who knows what.
To test what that directory is worth, Wegner, Ralph Erber and Paula Raymond brought in 118 people who had been dating someone for at least three months. Some were paired with their own partner, others with someone else’s. Each pair studied items from everyday categories, and later each person was tested alone. With no instructions on who should remember what, real couples remembered more than the impromptu pairs. But when the researchers assigned categories (you take food, your partner takes history), the couples did worse than the strangers. The authors’ reading: the couples already had an arrangement, and the new rule cut across it.
In a 1995 study, Diane Liang, Richard Moreland and Linda Argote taught people to assemble transistor radios, some in groups and some alone. A week later, groups whose members had trained together recalled more of the procedure and built better radios than groups put together from people trained alone. The authors concluded that group training helped mainly by building a transactive memory system. Kyle Lewis later turned the idea into a 15-item questionnaire, tested in the lab and in the field. As the psychologists Julie Hewitt and Lynne Roberts summarize it, the scale measures three things: specialization, credibility and coordination.
Much of this evidence comes from labs and simple memory tasks; Lewis built his scale because, as he wrote, field research “has lagged.” What the studies do suggest is that a directory takes time to build, and that changing it carries a cost.
In Ohio, part of the operators’ directory was a machine. When something trips, the alarm will say so. The task force found that FirstEnergy “lacked procedures to ensure that its operators were continually aware of the functional state of their critical monitoring tools,” and that its “computer support staff and operations staff did not have effective internal communications procedures.” Those were two findings among many; the task force also faulted the company’s management of tree growth, its grasp of its own system, and the regional coordinators’ real-time support.
The client who pays on one condition
Imagine you take over a five-person finance team. The analyst who has handled your biggest client’s invoices for six years moves to another department. Nobody wrote down that this client pays only when the purchase order number sits on the first line. Everyone just knew to ask her.
Three months later, invoices start coming back. She’s two floors up, but nobody connects the problem to her, because nobody knew it was hers. A twenty-minute talk before she moved, built around one question (what do people usually come to you for?), would have put that rule on a list.
A directory for agents
AI agents are getting the same arrangement. Under the Agent2Agent protocol, which Google announced in April 2025, an agent publishes an “Agent Card,” which the specification defines as a document “describing its identity, capabilities, skills, service endpoint, and authentication requirements.” The stated aim is for agents to work together “without needing access to each other’s internal state, memory, or tools.” No shared head, just a shared idea of who can do what.
That directory can go stale the way a person’s does. The specification says card content “changes infrequently” and recommends standard web caching. Consider a scenario. An orchestrating agent sends refund requests to an agent whose card lists a refunds skill. The team behind that agent narrows it to one currency and forgets to update the card. Nothing breaks loudly. Refunds in other currencies just stop going through.
You can test for that. Keep known cases for each skill a card claims, and rerun them whenever the agent behind it changes. Log which agent got each task, so you can see the directory in use. And check that an agent says “I can’t do this” out loud, instead of returning nothing.
In Ohio, the IT staff learned the alarms were dead at 3:42, from an operator. The blackout began less than half an hour later and reached an area of an estimated 50 million people.