Spatial AI, in plain words.

Spatial AI is artificial intelligence that reasons about real places: where things are, how they change over time, and what is true on the ground right now.

• ask“how green is Cubbon Park?” and get the newest signed NDVI back, with its date and satellite · decoded on the home page
• answerabout a specific field or address, because the answer comes from observations of that place, not from text about it

Earth memory is a record of what has been observed at each place on Earth, stored so an AI agent can recall it later and prove where it came from.

• keepeach observation at a ~10 m address as a small signed fact: its value, its time, its source
• seeCubbon Park, every layer: 26 signed facts from 11 layers at one place, bound into one handle

Grounding is giving an AI agent a checkable source for what it says about the world, instead of letting it answer from memory or guesswork.

• citea signed observation and its date, not a sentence that sounds right
• catcha draft that says Cubbon Park’s NDVI is 0.81 while citing its token: denied, PROV_VALUE, fix correct_value · try it

Each fact carries an ed25519 signature and a blake3 content hash. Anyone can recompute the hash and check the signature offline, without trusting the server that served it.

• hashthe fact’s canonical bytes: blake3(GET /v1/facts/eheadieomxr2…) returns eheadieomxr2…, its own name
• verifythe ed25519 receipt offline, against the responder’s key; change the bytes or the address and the check fails · watch it

Imagery and maps show a place to a person. Earth memory gives an agent a signed answer about that place that it can cite.

• compareSentinel-2B’s 351.0 MB scene over Bengaluru with its emem note: an agent reads ~4.9k context tokens instead of ~223M, 45,510× less · run it
• readwhat changed and whether a reading can be trusted, which a picture or a map does not say