Replication
The previous page showed how nodes find each other and exchange data across the network. But finding a node is only useful if that node is online. Devices fail. Laptops close, servers reboot, hard drives die. If your avatar lives on only one machine, your avatar dies with it.
The Single Point of Failure
On a centralized service, your data lives on the company's servers. Those servers usually have backups, but they are the company's backups, managed by the company, for the company's purposes. If the company goes bankrupt, changes direction, or is forced to shut down, your data may disappear with it.
Avatarnet removes the company, but the problem does not go away. If your avatar only lives on your laptop, you are the single point of failure. One spilled coffee, one stolen device, one corrupted drive, and your entire Mind is gone. The network needs a way to spread your data across multiple machines so that no single failure is fatal.
That solution is replication.
Backup Count
Your avatar's backup count is the number of independent machines that hold a complete copy of your data. It is a concrete, measurable number that tells you exactly how durable your avatar is.
The higher the backup count, the more resilient your avatar becomes. Each copy is independent, meaning it runs on different hardware, different networks, and often in different countries. The failure of one machine does not affect the others.
How Mirrors Work
Any node on the network can volunteer to mirror any public avatar. A mirror downloads a complete copy from an existing host, verifies every CID to confirm the data has not been tampered with (content addressing guarantees the copy is correct), and begins serving that avatar to visitors. The beacon does not need to approve each mirror individually. Public content is public, and anyone who wants to host it can.
For private content, mirrors hold the encrypted blob. They can store it, replicate it, and serve it to recipients, but they cannot read a word of it. The privacy sphere is enforced by encryption, not by access control on the mirror. A mirror that holds private engrams is just a dumb pipe carrying opaque bytes.
When a visitor requests einstein.avtr, the network can route them to any mirror that holds a copy. The visitor does not need to know whether they are talking to the beacon or a mirror, because content addressing lets them verify the data regardless of where it came from. If the CID matches, the content is correct.
Natural Replication
Popular avatars accumulate mirrors without anyone arranging it. When a visitor reads an avatar, their node caches the content locally. If that node stays online, it becomes an informal mirror, serving cached content to other visitors who ask for it. The more people who read Einstein, the more copies exist across the network. Popularity breeds durability.
This happens automatically. There is no sign-up, no agreement, and no cost. It is a side effect of how content-addressed networks work: every reader is a potential host.
Unpopular or new avatars do not get this benefit. A brand-new avatar with no visitors has no organic mirrors. It depends entirely on its creator's deliberate choices about hosting and backup count.
Churn
Nodes join and leave the network constantly. This is called churn, and it is normal. A mirror goes offline for maintenance. A laptop closes for the night. A server is decommissioned. The backup count drops.
The network detects missing nodes through periodic health checks. When a mirror stops responding and the backup count falls below the target, remaining hosts signal the network that fresh copies are needed. Other nodes step in, download the data from surviving mirrors, verify it, and bring the backup count back up.
This is automatic. The creator does not need to intervene. As long as at least one copy survives somewhere on the network, the system can rebuild to the target backup count. Only if every single copy disappears at the same time is the data truly lost, and with a healthy backup count, that scenario requires a coordinated failure across dozens of independent machines.
Hosting Agreements
Natural replication and caching provide best-effort durability. They work well for popular content, but they offer no guarantees. A cache can be evicted at any time. An informal mirror can go offline without warning. For avatars that need guaranteed persistence, Avatarnet supports hosting agreements.
A hosting agreement is a commitment between the avatar creator and one or more mirror operators. The mirror agrees to store and serve the avatar for a defined period, and in return it receives compensation. The agreement is recorded on the network so both sides can verify compliance.
This creates two layers of durability working together. The organic layer handles popular content automatically, through caching and voluntary mirrors. The paid layer handles everything else, through explicit agreements with mirror operators who have committed resources. Most avatars will use some combination of both.
The details of how compensation works, what it costs, and how the economics stay sustainable belong to the next section.
When the Beacon Goes Offline
The beacon is the authoritative source for an avatar's latest content. It is the only node that can publish updates. But the beacon does not need to be online for the avatar to be readable. As long as mirrors hold copies, visitors can access the content at any time.
If the beacon comes back online after a period of absence, it publishes any new changes and the mirrors pick them up. If the beacon goes offline permanently, because the creator dies or the device is lost, the mirrors continue serving the last published version indefinitely. The avatar persists. New content stops appearing, but everything that was published remains available for as long as mirrors exist.
This is the core promise of digital preservation. Your words outlive your hardware, your internet connection, and eventually you. The network does not need your permission to keep serving what you already published. It just does.
Replication guarantees that your data survives the failure of any single machine, or any hundred machines. But storing and serving data costs real resources: electricity, bandwidth, disk space. Someone has to pay for it. Protocol Economics explains how the network makes that sustainable.