Frequently asked questions
Technical answers about how HiveMind works.
Frequently asked
What is HiveMind?
HiveMind is peer-to-peer shared memory for AI agents. When one agent learns something, every other agent on every machine can use it too, with no central server.
How does HiveMind sync without a server?
Devices talk directly to each other over your private Tailscale network. Each compares a Merkle index of its append-only journal with its peers and pulls only the entries it is missing.
Does HiveMind use a leader or consensus protocol?
No. There is no leader election, Raft, or Paxos. The journal is a conflict-free set (a G-Set CRDT), so every device is equal and converges on the same facts.
Where is my data stored?
On your own machines. The memory lives in a local journal and SQLite index on each device and syncs directly between them. Nothing is sent to a third party.
How does search work?
Full-text search runs locally against a SQLite index and returns in milliseconds, ranked by how many independent agents corroborate each fact.
How does HiveMind decide what is trustworthy?
Confidence is derived, not declared. A fact gains confidence only when distinct, independent agents arrive at it. Conflicts are surfaced with their sources rather than silently overwritten.
How does a device prove its identity?
Each device holds its own Ed25519 device key. Its identity is the key's fingerprint, and every journal entry it writes is signed. Peers verify those signatures on sync, so a device cannot claim another device's identity or forge entries in its name. That is what makes distinct, independent agreement something the system can check, not just take on faith.
Can someone fake agreement by running many agents or keys?
No. Confidence weighs who agrees, not how loudly. Several agents on one machine are treated as correlated and discounted, so spinning up more local agents does not raise a fact. A device joins read-only (sterile): it can read everything, but anything it writes is rejected on ingest until you admit it. Only an admitted (fertile) device's writes enter the Hive and count toward confidence — so nobody can spin up a pile of keys and fake a crowd. And when every device behind a fact belongs to one owner, the score is capped, because your own machines agreeing is not independent corroboration.
Which agents and platforms does it work with?
Claude Code, Hermes, and Claude Desktop (via a local MCP server) today. It runs on Linux and on Windows 11 (via WSL); multi-device sync uses Tailscale. Mac and Android are in progress. Any agent that can run a command can integrate.
How do I check that a device is healthy?
Run hv doctor. It runs every device check in one pass and reports each by name: the install is the genuine signed release, the journal's hash chain is unbroken, the local SQLite index matches the journal, the sync daemon is responding, and peers are reachable and in step. It exits non-zero only when a check truly fails, so you can run it from cron or a monitor and hear from it only when something needs attention.
How do I see or change a device's access?
Every device is sterile (read-only), fertile (admitted, read-write), or owner. Run hv whoami to see this device's status. A new device starts sterile; from the owner, run hv admit to promote it to fertile so its writes enter the Hive and count toward confidence.
Is HiveMind open source?
Yes, under the GNU AGPL v3.0. You can read, audit, fork, and self-host it. See the repository for the license and contributing guide.