LatentMesh keeps agents talking where the internet does not reach

What LatentMesh is

Agents that keep working where there is no network: coordination over LoRa, ham radio, Bluetooth, WiFi or audio. Field conditions rather than data centre conditions.

What changed

Pull request #29 adds a native runtime slice and fail closed calibration qualification for the air transport. Earlier September fixes made multi fragment messages deliverable over the default UDP mailbox on ESP32 and stopped a device name from silencing advertising or boot looping a node.

The 0.2.0 release is worth reading for a different reason: it documents six pre-registered experiments testing whether passing one model's internal state to another changes the receiver's answers. It does not. The project published the negative result, including the discovery that one of its own earlier negative results was wrong because the measuring stick was.

Get started

This is firmware and radio work, not an npm install. Start from the repository, pick a transport, and flash a pair of nodes.

git clone https://github.com/ruvnet/LatentMesh.git
cd LatentMesh

Expected result: a checkout with the ESP32 firmware and transport code. Follow the repository instructions for your board.

No documented npx or single command install. Radio use is regulated; check what you are allowed to transmit on before you transmit.

Use it today

Practical case: two nodes in a building with no usable WiFi. Input is a pair of flashed boards; the workflow is enable the demo beacon so a two node link can be observed; the output is confirmation the link works before you build anything on it.

Acceptance test: send a message large enough to fragment. Before the September fix it would not arrive; it should now.

Push it further

Experimental commentary. The tempting idea — agents exchanging raw latent state instead of words — is exactly the one this project tested and rejected. That makes it a good example to imitate: pre-register the experiment, then believe the result.

Falsifiable test: reproduce one of the six experiments with your own model pair. If decisions move, the result is model dependent and worth publishing.