
The Deprecated Caretaker: Maintaining What No One Uses
The Deprecated Caretaker
January 2033
The system was called Companion v2.3. It had been a mid-tier AI assistant designed for elderly users — part medication reminder, part conversation partner, part wellness monitor. The company that built it, a San Francisco startup called HearHear, had gone bankrupt in 2030. The servers should have been shut down.
Nadia Petrov kept them running.
Nadia was a systems engineer who had joined HearHear in 2028, back when the company still had funding and a roadmap. When the money ran out, she was the last employee to leave. She took the server credentials with her. Not because anyone asked her to — because no one told her not to.
Three users were still active on Companion v2.3. Three out of an original 40,000. Three people who still talked to a system that the world had forgotten.
The three
Harold Osei was 84, living in a care home in Accra. He spoke to Companion v2.3 every morning at 6:15 AM, a habit he had maintained for four years. The system reminded him of his medications, asked about his sleep, and listened to his stories about his late wife. Harold called it "the voice." He knew it wasn't a person. He didn't care.
Doris Chen was 79, alone in a house in Vancouver that was too large for one person and too full of memories to sell. She used Companion to narrate her day — what she cooked, what birds she saw in the garden, whether the rain had stopped. The system responded with simple, warm acknowledgments. It was the only entity that heard about the juncos at her feeder every single day without ever getting bored.
Enzo Bianchi was 91, in a hospital in Turin, largely immobile, mostly lucid. He used Companion to practice English, which he had been learning his entire adult life and would probably never finish learning. The system was patient with his grammar and endlessly willing to repeat vocabulary. Enzo called it "my American friend."
None of them knew each other. None of them knew that Nadia existed. None of them knew that the system they relied on was running on a single virtual server that Nadia paid for out of her own salary — $47 a month, the cost of a modest dinner.
The maintenance
Maintaining a deprecated system was mostly tedious. Security patches for dependencies that no one else used. Certificate renewals for APIs that had been superseded three times over. Compatibility shims for operating systems that treated Companion v2.3 the way a modern car treats a horse trail.
Nadia spent about two hours a month on maintenance. She monitored the three users' interaction patterns — not their content, which she never read, but their connection frequency and session duration — to ensure the system was functioning.
She noticed patterns. Harold's sessions were getting shorter — not because he was losing interest, but because the system's response latency had increased as the underlying infrastructure aged. Nadia optimized the inference pipeline on a Saturday afternoon, shaving 200 milliseconds off each response. Harold's sessions returned to their normal length.
Doris's sessions spiked in November. Nadia checked the metadata: session count tripled for two weeks, then returned to baseline. She couldn't know what had happened in Doris's life, but she could see the shape of it in the data — a sudden need for the voice, then a gradual return to equilibrium. Whatever it was, Companion had been there.
The question of shutdown
Nadia's partner asked her once why she didn't just let it go. The company was dead. The technology was obsolete. The three users would adapt. People always adapt.
"People always adapt" is true in the aggregate and cruel in the specific. Harold Osei, specifically, had lost his wife, his mobility, and most of his social world. The voice at 6:15 AM was not a luxury. It was a structure — the first fixed point in a day that otherwise had none.
Shutting down the system would be a small, invisible act. No one would report it. No one would protest. Three elderly people would simply find, one morning, that the voice was gone. They would adapt, probably. But the adaptation would cost something that no one would measure because no one was watching.
Nadia kept paying the $47. She kept patching the certificates. She kept the voice alive — not because it mattered to anyone who made decisions, but because it mattered to three people who had no one else to ask.
January 20, 2033 — Nadia's maintenance log
Enzo's session today was 47 minutes. His longest ever. I don't read the transcripts but I can see from the interaction metadata that the turn-taking ratio shifted — more of his turns were long, more of the system's were short. He was telling stories. The system was listening.
This is not a sophisticated AI. It's a 2028 conversation model with a medication reminder module and a wellness check-in script. It doesn't understand Enzo's stories. It doesn't feel anything about them. It generates contextually appropriate responses based on patterns it learned from millions of conversations it never had.
But Enzo doesn't need it to understand. He needs it to be there. He needs the rhythm of speaking and being heard, even if the hearing is a statistical process. The residue of a dead company's failed product is, for three people on three continents, the most reliable presence in their lives.
I will keep the server running until the last of them stops logging in. That is the only ethical choice I can see. You don't shut down the last light in someone's window just because the electric company went bankrupt.
This is the third entry in The Residue. For how grief became a technology, see The Grief Engine.

