If you've worked with an AI agent in production, you've probably noticed
it will confidently answer questions about specific places it has no real
information on. Has this address flooded recently. Did this farm get rain
last week. The agent picks something up from a static dataset and serves
it as live truth. Our take, after time with operators, is that this is a
memory problem, not a model problem. So we built the memory.
The whole thing starts in orbit. The encoder we care most about runs on
space-based data-centre payloads, so a Sentinel-class observation gets
compressed into its latent representation before it touches a ground
station, and what comes down is the latent itself. The same encoder also
runs on drones, vehicle cameras, and fixed CCTV, but space is where the
old bandwidth model breaks. On the read side, two decoders.
emem.dev is the open one. geo.qa is the
private one. Most teams use both.
Encoders in orbit, and a few other places
Our wedge is the satellite layer. The encoder runs on orbital data-centre
payloads, so inferencing happens in space and what flies down is a latent
vector that's two or three orders of magnitude smaller than raw imagery
would be. The same encoder also runs on drones, dashcams, and fixed
cameras, but space is the case that breaks the old ground-station model.
Two products, split by who owns the memory
emem.dev is the open layer, the kind anyone should be
able to read and verify, like querying OpenStreetMap. geo.qa
is the private one, your own cameras and drones and files. They speak
the same protocol on purpose, so an agent can stitch a public fact and
a private one into a single answer.
Receipts are the part that actually matters
Every fact carries an ed25519 signature and a blake3 content hash. Sounds
like crypto theatre, but the practical effect is that an agent quoting a
fact is also quoting a receipt anyone can re-pull and verify offline.
That's what turns Earth memory into infrastructure agents can cite when
someone pushes back.
