emem

Earth as memory, for real-world agents.

Hosted · Docs · Spec · OpenAPI · Try it · /verify · Gallery · HF Space

---

Where emem has attested facts right now. 1440×720 plate-carrée. The same SVG renders live at /v1/coverage_map.svg.

Mumbai · Manhattan · Tokyo, painted by Copernicus DEM elevation. Each image is a live endpoint URL; click to see the latest signed render. /docs/gallery has the full set.

---

An LLM asked what is at a place will usually guess. It has no stable handle for the patch of ground at 19.07° N, 72.87° E and no audit trail for the number it returns. emem is the missing handle. Every cell on Earth gets a 64-bit identifier (about 9.55 m on a side at the equator). Every measurement at that cell is recorded as a fact keyed by cell, band, and time, signed with ed25519 over the blake3 hash of its canonical CBOR. Paste the 26-character content id into a chat and your colleague can pull the same bytes from any responder and verify the signature in their browser at /verify. The hosted node is https://emem.dev; there are no keys, no accounts, and the same handlers answer both MCP and plain REST.

When an agent asks for a band at a cell that has no signed fact yet, the responder fetches the underlying tile through one of its forty-three upstream sources, signs the result as its own primary attestation, persists it, and returns in the same response. The cold path takes about 180 ms; warm reads are sub-ten milliseconds. Every cell on Earth answers cite-ably from day one, without a pre-seeded corpus.

Three foundation encoders sit GPU-pinned alongside the responder: Clay v1.5 reads Sentinel-2 at a 2.56 km receptive field, Prithvi-EO-2.0-300M-TL reads HLS at 6.7 km, and Tessera reads Sentinel-1 plus Sentinel-2 per-pixel. Each carries its own aliasing pattern, so disagreements are informative. The clay_prithvi_tessera_triple_consensus@1 recipe votes the three; six domain variants follow for deforestation, wetland change, urban expansion, disaster anomaly, climate archetype, and coastal erosion. Receipts pin the algorithm CID, so a third party can replay the score against the same input facts and reproduce the same number.

When a band genuinely has no data at a cell (the encoder is offline, the place is outside coverage, the archetype seed never materialised), the responder returns a signed absence with a typed reason. An empty answer is itself a citable receipt, not a 404 and not an empty array. Whitepaper at /whitepaper.md walks the math.

Try it (no install, no key)

# Geocode a place to a cell64.
curl -s -X POST https://emem.dev/v1/locate \
  -H 'content-type: application/json' \
  -d '{"q":"Bengaluru"}' | jq .cell64
# "defi.zb493.xoso.zcb6a"

# Recall a band at that cell (auto-fetched if cold).
curl -s -X POST https://emem.dev/v1/recall \
  -H 'content-type: application/json' \
  -d '{"cell":"defi.zb493.xoso.zcb6a","bands":["weather.temperature_2m"]}' \
  | jq '.facts[0]'

# Ask a free-text question; the foundation-embedding fan-out fires
# automatically on "find places like" / "what changed" intents.
curl -s -X POST https://emem.dev/v1/ask \
  -H 'content-type: application/json' \
  -d '{"q":"find places like Yellowstone","place":"Yellowstone National Park"}' \
  | jq '.foundation_embeddings'

# Hunter mode: discover event hotspots over a named region. The same
# classifier reads "find <event> in <region>" from /v1/ask and routes
# here; structured callers can hit /v1/hunt directly.
curl -s -X POST https://emem.dev/v1/hunt \
  -H 'content-type: application/json' \
  -d '{"event":"algal_bloom","region":"Lake Erie"}' \
  | jq '.hotspots[0]'

The receipt's fact_cid is a durable handle. Re-fetching it from any responder, in any year, returns the same bytes.

Verify an answer (four curls)

The pitch lives or dies on this flow. Every recall response carries a receipt with fact_cids[], a merkle_proof, and an Ed25519 signature over the canonical preimage blake3(request_id ‖ served_at ‖ primitive ‖ cells ‖ fact_cids) — UTF-8, sections joined by |, list elements followed by ,. The signer's public key is stable; the receipt verifies offline against any copy of the responder pubkey.

# 1. Resolve a place to a cell64.
CELL=$(curl -s -X POST https://emem.dev/v1/locate \
  -H 'content-type: application/json' \
  -d '{"q":"Golden Gate Park, San Francisco"}' | jq -r .cell64)

# 2. Recall a band and capture the receipt envelope.
curl -s -X POST https://emem.dev/v1/recall \
  -H 'content-type: application/json' \
  -d "{\"cell\":\"$CELL\",\"band\":\"indices.ndvi\"}" > /tmp/recall.json

jq '.receipt | {primitive, served_at, responder_pubkey_b32, fact_cids, merkle_proof: .merkle_proof.root}' \
  /tmp/recall.json

# 3. Ask the responder to verify its own signature (server-side check).
jq '{receipt: .receipt}' /tmp/recall.json > /tmp/receipt.json
curl -s -X POST https://emem.dev/v1/verify_receipt \
  -H 'content-type: application/json' --data @/tmp/receipt.json
# {"valid":true,"preimage_blake3_hex":"…","fact_cids_count":1,"signer_pubkey_b32":"…",…}

# 4. Reproduce: pull the same fact_cid from any responder, on any day.
# The cell, band, tslot, and derivation.fn_key are content-addressed —
# the bytes you receive will hash to the same fact_cid.
jq '.facts[0].derivation' /tmp/recall.json

For a browser-only verify, open /verify/<fact_cid> — the page does the same Ed25519 check in WebCrypto + @noble/ed25519 so you never have to trust the responder you got the receipt from. A guided walk lives at /demos/signed-answer.

Connect your AI assistant

The MCP endpoint is https://emem.dev/mcp. Drop a config snippet into your client.

Client	Config
Claude Desktop	examples/claude-desktop.json
Claude Code	examples/claude-code.mcp.json
Cursor	examples/cursor.mcp.json
Cline (VS Code)	examples/cline.mcp.json
Gemini CLI	gemini extensions install https://emem.dev/gemini-extension.json
ChatGPT (Custom GPT)	examples/openai-gpt-action.json
LangChain (Python)	examples/langchain.py
LangChain MCP agent	examples/langchain/
LlamaIndex (Python)	examples/llamaindex.py
LlamaIndex MCP agent	examples/llamaindex/
Agno MCP agent	examples/agno/
Pydantic AI MCP agent	examples/pydantic-ai/
AutoGen MCP agent	examples/autogen/
CrewAI MCP agent	examples/crewai/
Mastra MCP agent	examples/mastra/

Python and TypeScript SDKs live under sdks/ (publication to PyPI / NPM pending; install from the repo today).

Primitives

58 MCP tools, 82 documented REST paths under /v1/*, surfaced through /openapi.json. Every tool carries a when_to_use string written for LLM tool-selection, and four MCP behavioural annotations (readOnlyHint, destructiveHint, idempotentHint, openWorldHint).

• Locate: name or lat/lng → cell64. Five-layer cascade: wide-bbox table → embedded gazetteer → GeoNames cities-5000 (68 581 places, in-process) → sled cache → Photon → Nominatim. Polygon geometry from Overture divisions/division_area. District-level queries reroute through Overture when Nominatim returns a POI courthouse.
• Memory substrate (state vector + memory token): POST /v1/state returns a signed dense per-place embedding (view=encoder default 128-D, view=cube full 1792-D). POST /v1/state_multi fans out across geotessera + clay_v1 + prithvi_eo2 with a typed missing[] for unwired encoders. POST /v1/state_diff returns the residual vector + L2 + cosine between two vintages at one cell. POST /v1/memory_token composes a memt:<cell64>:<fact_cid> handle any pipeline can paste; POST /v1/memory_token/resolve deferences it in one round-trip.
• Recall / recall_many / recall_polygon: 118 materializer-wired band names across 35 cube slots. Auto-fetch on miss; signed Absence on out-of-coverage.
• Find similar: k-NN over any vector band. Hamming fast path (sign-bit pop-count) auto-derives from the cosine band when the binary sibling is absent. Mode hamming_then_rerank triages with Hamming then re-orders by cosine; the over-sampling factor is EWMA-adaptive.
• Compare / compare_bands / diff / trajectory: pairwise and time-series.
• Verify: structured claim against attested facts; returns signed verdict + evidence CIDs.
• Physics: /v1/heat_solve (2-D explicit FTCS heat, MODIS LST stencil), /v1/wave_solve (1-D shallow-water along seaward bathymetry gradient), /v1/jepa_predict (closed-form NDVI AR(2) seasonal), /v1/jepa_predict_v2 (Tessera embedding dynamics; short-circuits to last-vintage identity baseline while the trained head is pending, receipt carries untrained_baseline).
• Ask: free-text question with topic routing. The classifier covers three intent families: place-anchored topical questions (the topic router fan-out), foundation-embedding intents on find places like / what changed / deforestation / anomaly (cross-encoder consensus over Clay + Prithvi + Tessera), corpus-meta intents on where do you have data / how fresh is your corpus (redirect to coverage surfaces), and hunter-mode discovery on find <event> in <region> (routes to /v1/hunt).
• Hunter: POST /v1/hunt and MCP emem_hunt for open-world event discovery. Twelve event keywords — algal_bloom, deforestation, flood_extent, wildfire, urban_heat_island, methane_plume, landslide, drought, soil_salinity, crop_stress, water_turbidity, oil_slick — each maps to a registered detection algorithm. The responder samples up to 32 cells from the named region (8 for slow primary bands such as MODIS LST), recalls the algorithm's primary scalar plus any configured gate band (e.g. NDWI > 0 for water-mask events), and returns the top 8 hotspots with cell64, lat/lng, recalled value, gate value, fact CID, and a Sentinel-2 scene URL. A Tessera embedding rerank fires when at least three candidate cells have a geotessera vector available, re-ordering by cosine similarity to the cluster centroid. oil_slick returns status: not_yet_implemented with pointers at flood_extent_sar_threshold@1 and water_turbidity_red_band@1 instead of fabricating detections.
• EUDR Due Diligence Statement: POST /v1/eudr_dds and MCP emem_eudr_dds produce a signed Annex II-shaped DDS under Regulation (EU) 2023/1115. The per-cell algorithm eudr_compliance@1 implements Article 2(4) as written: >0.5 ha, >5 m height, >10 % canopy cover, excluding land predominantly under agricultural or urban use. Forest baseline is JRC GFC2020 V3 (the Commission's expected non-binding baseline; live single-COG materializer at JEODPP) confirmed against Hansen GFC v1.12 treecover2000. JRC TMF (annual change, deforestation year, degradation year, transition subtype) reads through a pull-and-cache connector — JEODPP's TMF endpoint serves 84 MB tiles without HTTP Range, so the responder fetches once into $EMEM_DATA/jrc_tmf_cache/ with atomic rename, then samples. The Article 2(28) dispatch picks POINT (≤4 ha non-cattle) vs POLYGON (>4 ha or any cattle plot under HS 0102/0201/0202). Sims et al. 2025 driver attribution and RADD Sentinel-1 alerts layer on as refinement: both currently materialize Absence with a structured NotImplemented reason because Zenodo and GFW S3 do not honour HTTP Range, and the responder will not fabricate a value for a connector it cannot read. Every response carries the explicit Article 9(1)(b) legality disclaimer — land tenure, FPIC, country-of-origin laws are structurally out of EO scope and need a partner module before TRACES NT submission. The JSON Schema at /v1/schemas/eudr_dds.json cites the exact EUR-Lex paragraph each field maps to.
• Domain shortcuts: emem_at, emem_ndvi, emem_air, emem_lst, emem_soil, emem_water, emem_forest, emem_weather. Collapse locate → recall → polygon-aggregate into one call by place name.
• Field boundaries: Fields of The World (~3.17 B field polygons, 241 countries, 10 m, CC-BY-4.0) via PMTiles range reads on source.coop.
• Visual surfaces: /v1/coverage_map.svg (1440×720 plate-carrée of attested cells, log-scale density) and /v1/places/scene_overlay.svg?place=…&band=… (per-place value-painted bbox grid; band-aware ColorBrewer ramps, horizontal legend, km scale bar, signed source line). The MCP equivalents return the same SVG as an EmbeddedResource block. The full set, plus the 32-diagram protocol/industry suite, lives at /docs/gallery and /docs/diagrams.

Algorithms

159 named composition recipes (flood_risk@2, walkability_score@1, heat_index@2, carbon_sink_score@1, eudr_compliance@1, forest_carbon_loss_co2_flux@1, enteric_ch4_dairy_tier1_ipcc2019@1, n2o_synthetic_fertilizer_ef1_ipcc2019@1, ...) live in a content-addressed registry. Each carries:

• formula: plain math the agent can read and apply.
• inputs: band keys with role + explanation.
• when_to_use: agent-targeted trigger guidance.
• citation: peer-reviewed source.
• accuracy_band: honest precision estimate, not marketing.
• parameters: typed tunable thresholds (gate, k, timeout, ...).
• learned_from: citation provenance for every tuned number. An auditor can trace any gate threshold back to a referee.

Algorithms with an evaluation: Expr AST are also re-executable in-process: the responder walks the AST against the snapshot recall and returns a signed composite scalar that any third party with matching algorithms_cid and input fact CIDs reproduces deterministically.

Browse at GET /v1/algorithms or per-key at GET /v1/algorithms/<key>.

Discovery

Designed for agents to read, not for humans to remember:

GET /openapi.json                  — OpenAPI 3.1 of every REST route
GET /v1/agent_card                 — live capability snapshot + manifest CIDs
GET /v1/tools                      — 58 MCP tools with when_to_use + annotations
GET /v1/algorithms?summary=true    — 159 algorithm keys + categories
GET /v1/topics                     — 27 topic-grouped bands + algorithms (router brain)
GET /v1/manifests                  — bands_cid, algorithms_cid, sources_cid, schema_cid
GET /v1/schemas/eudr_dds.json      — Annex II JSON Schema with EUR-Lex paragraph citations
GET /.well-known/{emem,agent,mcp,ai-plugin}.json
POST /v1/state                     — signed dense state vector at any cell (view=encoder | view=cube)
POST /v1/state_multi               — fan-out across geotessera + clay_v1 + prithvi_eo2 with typed missing[]
POST /v1/state_diff                — vintage delta at one cell: residual vector + L2 + cosine
POST /v1/memory_token              — compose memt:<cell64>:<fact_cid> citation handle
POST /v1/memory_token/resolve      — single round-trip dereference back to signed fact body
GET /v1/stream                     — Server-Sent Events corpus heartbeat, signed every 5-300 s
GET /v1/corpus_state_stats         — signed snapshot of corpus liveness (one-shot equivalent of /v1/stream)
GET /v1/benchmark                  — hand-verified eval items; pair with POST /v1/benchmark/grade
POST /v1/hunt                      — structured event-discovery sweep (12 events × region)
POST /v1/eudr_dds                  — EUDR Due Diligence Statement (Regulation EU 2023/1115)
POST /mcp                          — JSON-RPC 2.0 (Streamable HTTP)
GET /llms.txt    /llms-full.txt    — plaintext catalog for LLM ingestion
GET /humans      /humans.json      — interactive try-it surface + machine twin
GET /verify      /verify/<fact_cid>— in-browser ed25519 receipt verifier
GET /docs/gallery                  — live coverage map + hunter case studies + 32 diagrams
GET /docs/diagrams/                — 32 SVGs of protocol + industry deployments

The operator_attestation block in /.well-known/emem.json binds the running binary's BLAKE3 hash to its git_commit + build_timestamp and signs the triple under the responder's ed25519 key, so a verifier can confirm the live binary corresponds to the published source tree without trusting the operator.

Every receipt pins four content-addressed registry CIDs (bands_cid, algorithms_cid, sources_cid, schema_cid). A peer that recomputes a fact under matching CIDs produces the same bytes. A peer with drifted registries returns a different bands_cid on /health and the divergence is visible before any data flows.

Run it locally

cargo run --release --bin emem-server
# Or via container.
docker run -p 5051:5051 ghcr.io/vortx-ai/emem:latest

No required env vars. EMEM_BIND overrides the listener (default 0.0.0.0:5051). EMEM_DATA overrides the data directory (default ./var/emem; pass :memory: for ephemeral). For TLS, systemd, ACME on :443, and the HuggingFace Space wrapper, see docs/operators/operating.md.

Address algebra

field	bits	wire form	example
cell	64	four base-1024 bigrams, dot-sep	defi.zb493.xoso.zcb6a
tslot	64	base32-nopad-leb128, t. prefix	t.aaaaagy
cid	32 B BLAKE3	base32-nopad-lowercase, 26 chars	qi3jo4sqcg…l2hgjtwm
vec	1792-D fp16	12-byte prefix in receipts	full vector via recall

The active grid is ~9.54 m × ~9.55 m at the equator (lat 21 bits × lng 22 bits, asymmetric to match the 360°/180° ratio). Above the equator, longitude pitch narrows with cos(lat). The Hilbert-ordered base-1024 alphabet keeps adjacent cells string-prefix-similar, so an LLM that emits defi.zb493… already lands in roughly the right place. GET /v1/grid_info declares the active resolution honestly; the spec target is a hierarchical migration toward H3-equivalent res-13 (~3.4 m).

Repo layout

emem/
├── crates/                       # 14 workspace crates, MSRV 1.88, version 0.0.6
│   ├── emem-core/                # bands, algorithms, functions, sources, topics, schema
│   ├── emem-codec/               # cell64, cid64, vec64, hilbert, geo, alphabet
│   ├── emem-fact/                # canonical CBOR; fact, receipt, attestation
│   ├── emem-claim/               # claim predicates (Op enum)
│   ├── emem-cache/               # sled cache wrapper
│   ├── emem-fetch/               # 18 data connectors + 5 utility modules
│   ├── emem-storage/             # sled hot cache + append-only merkle log
│   ├── emem-cubes/               # 1792-D voxel cube handle
│   ├── emem-primitives/          # recall, find_similar, trajectory, compare, diff, verify, query_region
│   ├── emem-attest/              # merkle root over fact CIDs
│   ├── emem-intent/              # rule-based intent → plan planner
│   ├── emem-mcp/                 # 50-tool MCP descriptor registry
│   ├── emem-api-rest/            # axum router, physics solvers, foundation fan-out
│   └── emem-cli/                 # binaries: emem-server, emem-livedemo, emem-realdemo, emem-demo, emem-ask-eval
├── sdks/
│   ├── emem-py/                  # Python client (httpx, sync + async)
│   └── emem-ts/                  # TypeScript client (zero runtime deps, native fetch)
├── python/                       # FastAPI sidecar over UDS: Prithvi-EO-2.0, Galileo, Clay v1.5, JEPA-v2
├── examples/                     # MCP configs + LangChain / LlamaIndex
├── ops/                          # systemd units, journald retention
└── web/                          # SSR HTML, humans, verify, llms.txt, agent.json

The 18 data connectors back 43 declared source schemes and 118 live materializer registrations. Most schemes route through cog.rs, the universal STAC + COG sampler, plus bespoke modules for chirps (rainfall), dmsp_ols (nightlights), esa_cci_biomass (above-ground biomass, CEDA), firms (active fire), ftw (Fields of The World), geonames (gazetteer), gmrt (topobathymetry, PointServer + GridServer), hansen_gfc (forest change), jrc_gfc2020 (EUDR forest baseline, JEODPP single-COG), jrc_tmf (tropical moist forest, pull-and-cache), koppen (climate classification), overture (places / buildings / divisions), radd_alerts (Sentinel-1 disturbance), terraclimate (climate), wdpa (protected areas), worldpop (population), wri_gdm_drivers (Sims et al. 2025 driver attribution).

Inference

The GPU sidecar (Python FastAPI over Unix domain socket) co-resides four encoders on a 20 GB VRAM budget:

• Clay v1.5: 1024-D CLS, S2 L2A 10 bands, ~12 ms warm. Teacher (DINOv2 vit_large_patch14_reg4_dinov2.lvd142m) pre-staged at boot so HF_HUB_OFFLINE=1 holds.
• Prithvi-EO-2.0-300M-TL: 1024-D CLS, HLS V2 6-band, ~13 ms warm.
• Galileo (variant base in production; tiny / nano selectable via EMEM_GALILEO_VARIANT): S2-only modality wired (S1 / ERA5 / SRTM / VIIRS / Dynamic-World / WorldCover / LandScan / location zero-masked; the scaffold is multimodal but only S2 is connected today). The advertised capability is galileo-<variant> in /v1/capabilities.extensions[].
• JEPA v2 dynamics: untrained baseline. Metadata-only is_trained() check short-circuits to last-vintage identity; receipt carries untrained_baseline and via: "short_circuit_untrained". Training is upstream-bottlenecked on multi-vintage Tessera availability.

Sidecar crash does not cascade. The REST router degrades to scalar bands and signs the GPU-anchored algorithms as Absence with gpu_unavailable. See docs/developers/inference.md.

Honest limits

• No commercial sub-meter imagery. Sentinel-2 (10 m), Landsat (30 m), HLS. For Planet Pelican (50 cm) or Maxar bring your own connector.
• No edge / onboard inference. Sidecar runs on a single host.
• Single-host deployment. No federation, no global routing, no SOC 2.
• JEPA v2 is untrained today. The endpoint exists and signs honestly; predictions equal the last attested vintage until the dynamics head is trained.
• 18 data connectors, 118 live materializer registrations. Catalog-by-count is not the pitch; every wired band is auto-fetchable, signed, and content-addressed. Bands without a wired materializer are listed under declared_but_no_materializer_at_this_responder.
• Foundation-encoder materializers are uneven. geotessera (Tessera 128-D) has a wired materializer and auto-fetches on miss. clay_v1 and prithvi_eo2 are seed-only at this responder — the GPU sidecar runs both models, but the auto-materialise path that fans out to upstream tile archives is not wired today. Recall against either returns whatever has already been signed; the hunter-mode envelope discloses this per request under materializer_status[].
• Tessera is upstream-rate-limited. dl2.geotessera.org reliably serves 2024 vintages today; historical backfill across all eight vintages (2017–2024) is partial. The Tessera-coherence rerank in hunter mode gracefully degrades to primary-scalar order when the upstream is unreachable, surfacing the reason under embedding_rerank.reason.
• MODIS LST is rate-limited. modis.lst_day_8day materialises through the NASA/ORNL REST API at roughly 30 s per cell. Hunter mode caps the per-region fan-out for the LST family to 8 cells (env override EMEM_HUNTER_SLOW_BAND_CAP) so urban-heat queries return inside the gateway timeout.
• No interactive notebook UI. For exploration there is /humans (try-it drawer, manifest grid, ontology SVG); for analytics, drive from a notebook against the REST or MCP endpoint.

Resources

Agent loop	https://emem.dev/agents.md
Wire spec	https://emem.dev/spec.md
llms.txt	https://emem.dev/llms.txt
OpenAPI 3.1	https://emem.dev/openapi.json
MCP	https://emem.dev/mcp
Verify	https://emem.dev/verify
Container	ghcr.io/vortx-ai/emem:latest (multi-arch, anonymously pullable)
HF Space	huggingface.co/spaces/vortx-ai/emem
MCP Directory	docs/mcp-directory.md
Issues / PRs	github.com/Vortx-AI/emem/issues
Security	SECURITY.md, avijeet@vortx.ai

License

Apache-2.0. See LICENSE and NOTICE.

Default-build data sources are open: Copernicus DEM, JRC GSW (CC-BY 4.0), Hansen GFC, ESA WorldCover (CC-BY 4.0), Overture Maps (places, buildings, transportation, divisions/division_area; ODbL / CDLA-Permissive), Fields of The World (CC-BY 4.0), GeoNames cities-5000 (CC-BY 4.0), OSM (ODbL), met.no, Open-Meteo, Tessera. No API keys, no operator credentials, no SaaS lock-in.