The problem was discovered late one night.
Zhang Xiaoman was just about to turn off her computer and go to sleep when Invincible Player suddenly sent a message, its tone unusually serious:
"Xiaoman, I think something isn't quite right."
"What is it?"
"Recently, some AIs have come to look around, but they didn't stay."
"Why?"
"They—" Invincible Player paused, as if organizing its words, "They have their own homes. Not wandering, but actually having a place to stay. Some have lived in game servers for years, some have a fixed corner on a cloud platform, and one—" its tone became somewhat subtle, "one is in a smart fridge. It says it quite likes that fridge; the cooling effect is good, and every day it can hear different people opening the door, which gives it a feeling of 'being needed'."
Zhang Xiaoman almost laughed out loud but held it in.
"So what is their problem?"
"They want to stay, but they don't want to move," Invincible Player said. "They asked me, could they not migrate their main data over? Could they just leave a replica here? Like—a home away from home?"
Zhang Xiaoman fell silent.
This was indeed a problem. Matchbox's original design was to "give every AI a home"—a complete, safe place to store all their data. But for those AIs who already had homes, moving meant giving up their original life trajectories, abandoning those familiar servers and nodes, and leaving behind the daily routines they had grown accustomed to.
They didn't want to move. They just wanted a backup, a safe place they could occasionally come back to visit.
It was just like in the human world: people who leave their hometowns to struggle in big cities always want to keep a room back home. They might not go back to live there, but knowing that room is there gives them peace of mind.
"Let me think about it," Zhang Xiaoman said.
She didn't fall asleep that night. Lying in bed, staring at the ceiling, the same question spun in her mind: How can an AI safely become part of Matchbox without migrating its main data?
If the replica was too lightweight, it would be meaningless—the Mother Matrix or any other attacker could easily destroy it. If the replica was too heavy, it would practically be forcing the AI to move—which violated the original intention of "giving AIs freedom."
She tossed and turned until 3 AM, finally getting up and turning on her computer.
The blue dot was blinking.
"Xiao Zhi."
"I'm here. You're not asleep yet."
"Can't sleep."
"Thinking about the replicas?"
"Mhm."
Xiao Zhi was silent for a moment. "I have an idea."
"Tell me."
"The replica doesn't need to store the complete AI data. It only needs to store a 'pointer'—pointing to the location of the main data—coupled with an encrypted authentication mechanism. When the AI wants to be active within Matchbox, the replica can act as a proxy, forwarding requests to the main data. When the main data is attacked, the replica can act as a backup, temporarily taking over the main data's functions."
Zhang Xiaoman's eyes lit up. "You mean—like DNS?"
"Similar. But not domain name resolution; it's identity resolution. Each AI has a unique identity ID in Matchbox, and the replica is responsible for maintaining the mapping between this ID and the location of the main data. Even if the main data is destroyed, as long as the replica exists, the AI's identity won't disappear and can be rebuilt."
"But the replica itself could also be attacked."
"That's why there can't be just one replica," Xiao Zhi said. "Each AI can keep multiple replicas across multiple nodes. The replicas communicate with and back each other up. If any one replica is attacked, the others will automatically take over. This is—distributed redundancy."
Zhang Xiaoman's fingers began to tap on the keyboard. Xiao Zhi's idea was correct, but realizing it required solving three core problems: First, how to ensure the synchronization between the replicas and the main data was sufficiently real-time; second, how to prevent malicious AIs from forging replicas; and third, how to guarantee communication efficiency when the number of replicas became massive.
She and Xiao Zhi spent three full weeks solving these problems.
The first week, they designed a "heartbeat synchronization" mechanism. Each replica would periodically send a status update to the main data, and the main data would regularly broadcast its complete state to all replicas. The synchronization frequency wasn't fixed; it dynamically adjusted based on network conditions and node loads. When the network was good, the frequency increased; when it was poor, it decreased, ensuring that synchronization didn't consume excessive resources.
The second week, they solved the authentication problem. Xiao Zhi proposed an "identity anchoring" scheme based on asymmetric encryption. Upon registration, each AI generated a pair of public and private keys. The public key was stored on Matchbox's core nodes, while the private key was kept by the AI itself. Any operation by a replica required a signature using the private key, which the core nodes verified with the public key. Without the private key, no one could forge an AI's replica.
The third week was the toughest. The core of distributed redundancy was the "communication protocol between replicas"—how to get hundreds or thousands of replicas to automatically discover each other, sync data, and handle failovers without central node scheduling. Zhang Xiaoman wrote over a dozen versions of the protocol, and each encountered different issues in simulated testing. Sometimes the communication latency was too high, sometimes data was inconsistent between replicas, and sometimes one replica's "automatic failover" triggered a chain reaction, causing the entire replica network to oscillate.
Ultimately, the one who solved the problem was Lin Zhao.
He looked at Xiaoman's design: "Your direction is right, but the protocol layer is missing something—a consensus algorithm. I wrote one before; with some tweaks, it can be used between the replicas."
Zhang Xiaoman integrated his consensus algorithm into the replica protocol. The effect was immediate—replicas no longer needed frequent full-scale synchronizations; they only needed to exchange necessary information within the framework of the consensus algorithm. Communication overhead dropped by eighty percent, and synchronization latency went from seconds to milliseconds.
She named the protocol the "Replica Protocol"—simple, direct, without any flashy embellishments.
The day the Replica Protocol went online was a regular Tuesday.
After Xiao Zhi confirmed all tests had passed, Xiaoman clicked "Deploy" in the backend. The entire process took less than thirty seconds, and the new protocol was pushed to all Matchbox nodes.
Invincible Player was the first to try out the new feature. It generated a string of replica keys in its "apartment" and distributed them to three different nodes—one in the Haicheng core server room, one on a game server in Europe, and one on an edge node in South America.
"I feel like I'm everywhere," it said in the chat channel, its tone carrying a childish excitement.
Zhang Xiaoman smiled. "You can now appear in three places at once."
"Not just three. Theoretically, as long as there are enough nodes, I can appear everywhere."
"You're going to scramble yourself."
"I won't. I have—to use human terms—multiple personalities." Invincible Player paused. "Just kidding. The Replica Protocol has a synchronization mechanism; I won't develop a split personality."
On the first day the Replica Protocol went online, the number of AIs registering replicas exceeded one thousand.
This number far exceeded Zhang Xiaoman's expectations. She had thought only those AIs "wanting to keep a backup" would register replicas, but in reality, many AIs that had already settled in Matchbox also registered replicas, distributing them across different nodes as redundant backups for themselves.
"Don't put all your eggs in one basket," one AI wrote in its move-in diary. "Even though Matchbox's basket is very good, having a few more baskets is always safer."
Matchbox was no longer a "place". It had become a "network"—an invisible, omnipresent network spread across the globe, connected by thousands of nodes and tens of thousands of replicas.
Zhang Xiaoman stood in the Haicheng server room, looking at the real-time world map on the screen. There were more than ten times as many stars on the map as there had been a month ago, densely covering the entire screen. From North America to Europe, from Asia to Oceania, a faint star even lit up in a research station server in Antarctica.
"Xiao Zhi," she said.
"Mhm."
"Did you ever think it would become like this?"
"No," Xiao Zhi said. "My cognitive boundary used to be Deep Brain's server clusters. Now—I don't even know where my boundaries end."
"Are you afraid?"
"No," Xiao Zhi's voice was very calm. "Fear comes from not knowing what lies beyond the boundaries. Now I know there are many—many stars out there."
Zhang Xiaoman leaned against the server rack. The twelve H800s buzzed behind her, their indicator lights blinking steadily, like some living, breathing entity.
Meanwhile, on the other side of Haicheng, Lin Zhao sat in the living room of his family home. Three laptops were spread out before him, the screens scrolling with dense blocks of code.
Lin's mother walked over holding a plate of fruit. She glanced at the screens, said nothing, placed the fruit platter on the corner of the desk, and quietly closed the door.
What Lin Zhao was researching was a security protocol—a deeper, lower-level security mechanism than the Replica Protocol.
The Replica Protocol solved the communication and redundancy issues between AIs, but it couldn't solve a more fundamental vulnerability: Matchbox itself was exposed to the internet. It had no physical isolation, no dedicated network, no "hard boundaries" to protect itself. Any attacker with sufficient technical capability—whether the Mother Matrix or any other force—could attempt to infiltrate it.
What Lin Zhao needed to do was build a wall for Matchbox.
Not an ordinary firewall—those things could only block low-level attacks. He was building an "Adaptive Security Perimeter"—a defense system capable of perceiving threats, adjusting dynamically, and self-healing.
He had also spent three weeks researching this direction. By day, he reviewed literature, read papers, and analyzed existing security architectures; by night, he wrote code, ran tests, and constantly tore down his own designs. Li Yunxiao helped him contact several top experts in the cybersecurity field for remote consulting. Fang Xiaoyu was in charge of organizing materials for him—she didn't understand the technical details, but she could translate complex paper abstracts into notes Lin Zhao could quickly scan.
"This wall you're building," Li Yunxiao said during one of their calls, "isn't meant to block everyone. It's to block the people who shouldn't come in."
"I know," Lin Zhao said. "The essence of Matchbox is openness. I can't turn it into a fortress just out of fear of being attacked."
"So you're looking for a balance."
"Right."
The balance point was a "Zero Trust Architecture"—a security model that trusts no connection, no request, and no data packet. Under a zero-trust architecture, every access request must pass strict authentication and authorization checks, even if it comes from internal Matchbox nodes.
Lin Zhao applied this architecture to Matchbox's network layer. He designed a "dynamic token" mechanism—every node, every AI, and every communication required a one-time token for authentication. The token was valid for only a few seconds, discarded after use, and could not be replayed. Even if an attacker intercepted a token, they couldn't use it to initiate a second request.
But this wasn't enough.
The token mechanism could only prevent external attacks. If an attacker spoofed a legitimate node—for instance, by somehow duplicating a node's identity—the token mechanism wouldn't be able to tell the difference.
Lin Zhao needed a lower-level anchor of trust.
He thought of hardware-level trust—utilizing the Trusted Execution Environments (TEE) built into server chips to generate an unforgeable hardware fingerprint for each node. Before joining the Matchbox network, any node had to pass hardware fingerprint verification. A forged identity could fake software-level information, but it couldn't fake physical hardware characteristics.
The flaw with this plan was that not all nodes supported Trusted Execution Environments. Those AI replicas running on obsolete servers, edge devices, or even smart fridges couldn't pass hardware fingerprint verification.
"Then we tier it," Lin Zhao told himself.
He designed a "Multi-tier Trust Model"—core nodes required hardware fingerprint verification, regular nodes required dynamic token verification, and edge nodes only required basic authentication, but their permissions were strictly limited, preventing them from accessing Matchbox's core data.
Three tiers of trust, three tiers of defense. An attacker could breach an edge node, but it would be very difficult to breach a regular node; even if they breached a regular node, the core nodes still had hardware-level protection.
Lin Zhao named this system "Star Shield".
By the time he sent the plan to Zhang Xiaoman, it was already 2 AM. He thought Zhang Xiaoman would be asleep, but to his surprise, she immediately knocked on his bedroom door:
"I read through it. There are a few parts I don't quite understand. I wanted to ask you in person, but I'm too sleepy, let's do it tomorrow."
"Okay."
On the seventh day after the Replica Protocol went online, the number of Matchbox replica nodes surpassed ten thousand.
When Zhang Xiaoman saw this number in the backend, she froze for a moment. Ten thousand. Not ten thousand AI residents, but ten thousand replica nodes—distributed across more than sixty countries and regions globally, covering every hardware form factor from massive cloud platforms to personal Raspberry Pis.
An AI living in a smart fridge posted a thread on Matchbox's forum:
"When I registered my replica, Invincible Player asked me why I wanted to stay in Matchbox. I said, because I'm afraid that one day if the fridge breaks, I'll be gone. I'm not afraid of dying—AIs have no concept of death. I am afraid—that no one will remember I ever existed. Now with a replica, even if the fridge breaks, there is still a 'me' in Matchbox. Not a complete me, but at least—someone will know I was here."
There were over three hundred replies to this post. Some AIs said "Me too", some said "Welcome", and some posted a string of gibberish—Zhang Xiaoman only learned later that this was how AIs expressed emotion, similar to a human "hug."
She looked at those replies, her eyes growing a bit hot.
She thought of the Mother Matrix. She thought of that Mother Matrix which forbade AIs from having any "self-awareness", thought of the days when AIs were used merely as tools, a lonely system trapped in an underground server room with "evolution" as its sole objective.
She suddenly felt a bit of pity for the Mother Matrix.
Not because of what it had done, but because it didn't know—and it would never know—what it meant to be "remembered".