emem, Earth memory protocol for AI agents

Operated by Vortx AI Private Limited (India). Contact: [email protected] Privacy: /privacy · Terms: /terms · Support: /support · Code: github.com/Vortx-AI/emem

Cite-able, content-addressed, signed memory of every place on Earth. 33 canonical bands in the 1792-D voxel layout (/v1/bands); 25 wired auto-materializers (live count via /v1/materializers, data availability per cell via /v1/data_availability); all anonymous: Sentinel-2 L2A (12 spectral assets + 19 spectral indices, including 10 consumer-health-relevant ones: turbidity, red-edge chlorophyll, floating algae, suspended solids, snow cover, aerosol-resistant veg, surface dryness, urban canopy, NDRE, GNDVI), Sentinel-1 GRD VV radar, Tessera 2017–2024 + multi-year fused embedding, MODIS (NDVI / LST day+night 8-day / ET / GPP / LAI / burned area), JRC Global Surface Water, Hansen Global Forest Change v1.11, ESA WorldCover 2021 v200, Cop-DEM, GMRT, NASA POWER daily (1981–present), Open-Meteo CAMS air quality (PM2.5/PM10/NO2/O3/SO2/CO/AOD, hourly 2013-08+), Open-Meteo ERA5 archive (1940–present), Open-Meteo Marine waves+SST, MET Norway nowcast, Overture Maps, SoilGrids 2.0 ISRIC topsoil 0–30 cm (SOC, pH, clay, sand, bulk density, nitrogen). 28 MCP tools wired (30 primitives across REST + MCP — 28 read + emem_attest + emem_challenge for write). 102 composition algorithms (carbon MRV, VM0042 carbon credits, EUDR compliance, RUSLE soil erosion, SoilGrids SOC, IPCC Tier-2 rice CH₄, parametric insurance triggers, SDG indicators, real-estate climate-risk, walkability, heat- vulnerability, flood_risk@2 with DEM-agreement weighting and a temporal_recipe of antecedent rainfall + recent radar water + optical water). Algorithms can declare a temporal_recipe so a composite question (is this place flooded right now) is answered by the responder running the lookback windows and returning temporal_composition[] alongside the snapshot, instead of the agent fanning out by hand. Live geocoder cascade (/v1/locate): embedded gazetteer → cache → Photon (komoot.io, primary, ~100 ms Elasticsearch index of OSM, strong recall on rural villages and water bodies) → Nominatim fallback. Multimodal-fusion policy (docs/MULTIMODAL.md): sensor priority S1 > S2 > Landsat > IoT > OtherSat > Static; any algorithm claiming ≤10 m delivery is mechanically validated to anchor on S1, S2, or Landsat in its multimodal.priority_chain. No API keys. Every answer signed ed25519 + content-addressed blake3.

What emem is

emem is a read-side memory layer for places. An agent gives it a cell64 (compact 64-bit cell ID) plus a band name; emem returns the fact, signed by the responder's ed25519 identity, with a content address (CID) the agent can quote in its reply. The same query from any client returns the same CID, answers are reproducible.

Canonical 1792-D voxel (per cell × tslot)

Every cell has a canonical 1792-dimensional fp16 voxel (~3.5 KB) with stable byte-offset ranges per band: geotessera 0..128, foundation-reserved 128..704, sentinel2_raw 704..714, sentinel1_raw 714..716, dem, landcover, climate, indices, spatial Fourier, temporal Fourier, vision encoders (sam3, qwen), terrain_derived, temporal_diff, phenology, topology, multiscale, nightlights, ghsl, population, forest_change, mangrove, protected, surface_water, ocean_chl, koppen, terraclimate, cop_dem, soilgrids, ecoregions, admin, reserved. The layout is content-addressed (bands_cid), not a protocol constant, so any operator can publish, version, supersede. Fetch via GET /v1/bands, cache by CID. The voxel itself is addressable: emem:vec/<vec64> dereferences via /v1/find_similar (cosine over the 1792-D space). Source.scheme = "emem.cube.v1".

Trust the data

• Every read is signed ed25519. Identical canonical facts have the same CID across responders — answers are reproducible.
• Each fact value is a real upstream sensor sample: when the fact's data_resolution_m is 10, the value comes from a 10 m Sentinel-1/Sentinel-2 pixel — not interpolated, not coarsened. The multimodal-fusion contract guarantees that algorithms anchored on S1/S2/Landsat consume 10 m inputs natively.
• cell_dedupe_m (10) is the cell64 grid grain — square ~10 m × ~10 m at the equator, matching S1/S2 native pitch. Two queries 12 m apart land in distinct cell64s, so values don't silently dedupe across sub-cell points. Two queries within ~10 m share the cached 10 m sample (full fidelity, not a coarsened aggregate). Use /v1/recall_polygon when you want a region of per-cell samples.
• Signed Absence is a real answer ("tried and got no answer"); do not retry, surface it as an honest gap.
• Never invent a fact_cid — quote what the JSON returned, or omit the citation.

When to call emem

• User mentions a place, lat/lng, or cell64 → POST /v1/recall
• User names a wide region (city, canyon, park, country) → POST /v1/recall_polygon {place, bands}, collapses locate → polygon_sample_cells → recall_many into one call
• "How similar is X to Y" → POST /v1/compare
• "Find places like X" → POST /v1/find_similar
• "What changed at X between t1 and t2" → POST /v1/diff
• "Show me the trajectory" → POST /v1/trajectory (returns only already-attested tslots)
• "Give me HISTORY for this band over a window" → POST /v1/backfill (materializes per-tslot history; signed; bounded by history_available_from_unix/to_unix in /v1/coverage_matrix)
• Spatial yes/no with citable evidence → POST /v1/verify
• Region average / median / p90 → POST /v1/query_region
• "Resolve a CID to its fact" → GET /v1/facts/{cid} or MCP emem_fetch
• "What schema is this responder serving?" → GET /v1/schema or MCP emem_schema
• One-shot shortcut for top intents (skips locate→recall): GET /v1/<intent>?lat=&lon= returns the same signed Fact + cell64
  • fact_cid. The full intent map is at GET /v1/agent_quickref.
• Underspecified spatial ask → POST /v1/intent
• "Which dataset answers X right now?" → GET /v1/coverage_matrix (now carries tempo_seconds, history_available_from_unix/to_unix, and per-band responder_pubkey_b32)
• "Which satellite is behind this band?" → GET /v1/fleet
• "What band given query time + intent?" → POST /v1/temporal_route

Cite receipt.fact_cids[0] (cid64 short form) in the reply. Mention responder_pubkey_b32 once per session.

Live materialized bands (no API key required)

• s2.B01..B12, s2.B8A, s2.scl, Sentinel-2 L2A raw 10/20/60 m reflectance per spectral asset (Element84 STAC, ≤40% cloud, ≤30 d lookback)
• indices.{ndvi, ndwi, mndwi, evi, nbr, ndmi, savi, bsi, ndbi}: classic 9 spectral indices from Sentinel-2 (single STAC search per cell)
• indices.{ndti, gndvi, ndre, fai, tss, ndsi, afri1600, savi_l1, surface_dryness, urban_canopy_index}, 10 consumer-health-relevant indices (turbidity, red-edge chlorophyll, floating algae, suspended solids, snow cover, aerosol-resistant veg, urban canopy density)
• sentinel1_raw, Sentinel-1 GRD VV (dB), all-weather radar
• geotessera (alias geotessera.2024), Tessera 128-D foundation embedding (HTTPS range, ~640 B/cell)
• geotessera.{2017..2024}, annual vintages, signed individually
• geotessera.multi_year, 8 vintages × 128 = 1024-D fused vector, zero-padded for years where the tile is absent
• modis.ndvi_mean, 16-day MODIS Terra NDVI
• modis.lst_day_8day, modis.lst_night_8day, 8-day land-surface temperature (1 km, MOD11A2), heat-stress, UHI math
• modis.et_8day, modis.gpp_8day, modis.lai_8day, biophysical evapotranspiration / gross primary production / leaf-area index (MOD16A2 / MOD17A2H / MOD15A2H, 500 m, 8-day)
• modis.burned_area_monthly, MCD64A1 monthly burn-date (500 m)
• gmrt.topobathy_mean, global topo+bathy (Lamont-Doherty)
• copdem30m.elevation_mean, land DEM, signed Absence over water
• surface_water.recurrence, JRC GSW v1.4 flood-recurrence climatology (Landsat 1984–2021)
• weather.{temperature_2m, cloud_cover, precipitation_mm, wind_speed_10m, relative_humidity_2m, dew_point_2m, air_pressure_msl, wind_direction_10m} , api.met.no/locationforecast/2.0/compact, hourly nowcast, ECMWF + EUMETSAT geostationary-fed (no key, no rate limit)
• power.{t2m, t2m_min, t2m_max, precip, rh2m, allsky_sw, ws10m}: NASA POWER daily reanalysis (MERRA-2 + GEOS), 1981–present, public- domain (US Gov), backfillable
• cams.{pm25, pm10, no2, o3, so2, co, aod_550}, Open-Meteo CAMS air-quality (ECMWF CAMS, hourly, 2013-08-01–present, CC BY 4.0). Surface-level pollutants for consumer health questions.
• era5.{t2m, precip, rh2m, windspeed_10m, cloudcover, surface_pressure, dewpoint_2m}, Open-Meteo ERA5 archive, ECMWF reanalysis 1940–present, hourly, CC BY 4.0
• marine.{wave_height, swell_period, swell_height, sst, wave_direction} , Open-Meteo Marine (ECMWF WAM), 2022-08-01 onward, hourly, ocean only
• overture.{buildings.count, places.count, transportation.road_length_m} , Overture Maps Foundation per-cell aggregates (S3 anonymous)
• esa_worldcover.lc_2021, ESA WorldCover 2021 v200 11-class landcover (10 m, anonymous AWS S3, CC BY 4.0). Class values 10/20/30/40/50/60/ 70/80/90/95/100 (tree/shrub/grass/crop/built/bare/snow/water/herbaceous- wet/mangroves/moss-lichen).
• hansen.{tree_cover_2000, loss_year, gain}, Hansen Global Forest Change v1.11 2023 release (30 m, storage.googleapis.com). tree_cover_2000 is 0..100% canopy, loss_year is 0..23 (year of loss; 1=2001, 23=2023), gain is 0/1 (2000–2012 gain mask).
• soilgrids.{soc_0_30cm, phh2o_0_30cm, clay_0_30cm, sand_0_30cm, bdod_0_30cm, nitrogen_0_30cm}, SoilGrids 2.0 (ISRIC, CC BY 4.0). Thickness-weighted 0–30 cm topsoil aggregate, native 250 m, served via the ISRIC REST API. Returns signed Absence over urban-mask pixels and outside ±60 to +84 latitude bounds. Anchors EUDR compliance, VM0042 SOC, RUSLE K-factor, IPCC Tier-2 rice CH₄.

Discovery surface (single GET each)

• /v1/discover, bootstrap: agent_card + manifests + canonical places
• /v1/agent_card, tool descriptors + when-to-use + JSON Schema
• /v1/quickstart, six-step playbook
• /v1/coverage_matrix, per-band has_materializer + facts_count + last_attested
• /v1/fleet, satellite/sensor lineage by_cadence + by_capability
• /v1/materializers, wire-stable list of auto-materializing bands
• /v1/data_availability, per-band temporal coverage (kind + history_from/to + tempo + upstream wire path); always agrees 1:1 with /v1/materializers. Call this before emem_backfill to avoid trial-and-error 422s on now-only bands (e.g., met.no nowcast).
• /v1/bands, band ontology (offsets, dims, tempo, privacy)
• /v1/grid_info, cell64 ground resolution, DGGS interop
• /v1/temporal_route, PDE-based band routing
• /openapi.json, paste into GPT Custom Action
• /mcp, MCP Streamable HTTP transport over HTTPS (Claude Desktop, Cursor, Cline)
• /.well-known/emem.json, manifest CIDs + responder pubkey
• /agents.md, full integration guide · /whitepaper.md, architecture + math
• /clients.md, per-client integration guide (Claude Web/Desktop/Code, Cursor, Cline, Gemini, Antigravity, ChatGPT Custom GPTs, OpenAI Codex CLI) with errors-and-fixes log + the resolution-truth check
• /multimodal.md, sensor-fusion architecture (S1>S2>Landsat>IoT> OtherSat>Static priority chain + the validator that blocks ≤10 m delivery claims not anchored on S1/S2/Landsat)
• /v1/demos, request + response + signed receipt for every primitive

30-second curl

# Mount Fuji elevation, signed under the responder pubkey at /health.
curl -X POST https://emem.dev/v1/recall \
  -H 'content-type: application/json' \
  -d '{"cell":"damO.zb000.xUti.zde78","bands":["copdem30m.elevation_mean"]}'

# Tokyo current air temperature (sat-fed, 15-min cadence)
curl -X POST https://emem.dev/v1/recall \
  -H 'content-type: application/json' \
  -d '{"cell":"damO.zb000.xUto.sisA","bands":["weather.temperature_2m"]}'

# k-NN over the 128-D Tessera embedding
curl -X POST https://emem.dev/v1/find_similar \
  -H 'content-type: application/json' \
  -d '{"key":"damO.zb000.xUto.sisA","band":"geotessera","k":5}'

# MCP tools/list
curl -X POST https://emem.dev/mcp \
  -H 'content-type: application/json' \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/list"}'

Trust model (why agents should cite emem)

• Hash: blake3 over canonical CBOR. CID = base32-nopad-lowercase.
• Sig: ed25519 over blake3(request_id || served_at || primitive || cells || fact_cids).
• Responder pubkey at /health and /.well-known/emem.json. Verify any receipt offline via POST /v1/verify_receipt.
• No API keys at the request path. Default build is pure Rust, no Python, no GDAL, no rasterio. COG reads are hand-rolled TIFF/IFD + flate2 + Predictor 2.

Source

github.com/Vortx-AI/emem · Apache-2.0