vortx-ai/emem/emem-locate-and-recall

emem-locate-and-recall

Network use. The commands in this skill send JSON to https://emem.dev and receive JSON from it. Where a step says so, they also receive an image or a raster file. https://emem.dev is the service that the MCP server of the plugin connects to. The commands do not execute what they download. This skill ships no scripts. The commands read no environment variable, key or credential. They read only local files that an earlier step saved from https://emem.dev.

Two calls change a place name into signed facts. Each fact comes back with its own emem:fact: token. Thus the number always travels with its citation.

When to use this skill

The user asks about a measurable geospatial value at a named place:

• "What is the current 2 m air temperature in Bengaluru?"
• "Show me the NDVI in the Sundarbans."
• "What is the elevation of Mount Kilimanjaro?"

Use a different path in these cases:

Step 1: resolve the place

Call emem_locate with this body:

{"q": "Bengaluru, India"}

Without MCP, send the same body to POST https://emem.dev/v1/locate.

Read cell64, place_label, via and disambiguation_required in the result. On 6 November 2026 this call returned defi.zb493.zezo.zcb35, "Bengaluru (city), Karnataka, India", through the photon geocoder. A repeat call for the same place can show cache in via. Other values are embedded, geonames, nominatim and direct.

The tool also accepts place as a synonym of q. A body with {"lat":..,"lng":..} does not use the geocoder.

Sometimes the name can mean more than one place. Then disambiguation_required is true, or selected.is_high_confidence is false. In that case, read alternatives and ask the user which place they mean. Do not cite rank 0. For example, "Springfield" returns Springfield, Missouri first, with disambiguation_required: true and five alternatives.

Step 2: recall the bands

Call emem_recall with the cell64 from step 1:

{"cell": "defi.zb493.zezo.zcb35", "bands": ["weather.temperature_2m", "indices.ndvi"]}

Without MCP, send the same body to POST https://emem.dev/v1/recall and save the result as recall.json.

Over MCP, save the JSON text of the tool result as recall.json. Then print the latest reading for each band, with the token to cite it:

jq -c '. as $r | .current_by_band | to_entries[] | .value as $c
       | $r.facts[] | select(.fact_cid == $c)
       | {band, value_verbatim, unit, observed_at, memory_token}' recall.json

Read the response by field name, not by position:

A timeout note with retryable: true means that the fetch continues in the background. The same call some seconds later usually answers from the store.

Two recorded runs at Bengaluru

The second run returned this temperature token: emem:fact:defi.zb493.zezo.zcb35:wmiie6af6m7ovchgs6ke36gak7fcowtz5zuroxwrwynlkbrlitba. The number of facts grows each time the responder stores a new reading.

How to choose bands

The band catalogue changes for each deployment. Ask the responder:

• GET /v1/bands lists the cube bands and the scalar keys of each band

(.bands[].key, .bands[].scalar_keys).

• GET /v1/materializers?page_size=100&summary=true lists the bands

that the responder can fetch on a miss. Follow pagination.next_url to get the next page. On 6 November 2026 the list on emem.dev had 113 entries on two pages.

These common scalars materialise automatically on emem.dev. All of them were in the materializer list on 6 November 2026.

emem.dev no longer computes the foundation-model embedding bands, for example clay_v1. Facts that these bands signed before still read and verify. A request for one of these bands returns a note with the reason band_retired_at_this_responder. Use the deterministic indices instead. Anyone can compute them again from the same Sentinel scene.

A value is a pixel sample, not a cell average

Each recall response has a spatial_basis block that says this clearly. The cell is an address, about 9.5 m across. The value is a sample of the source pixel that overlaps the cell. One 30 m DEM pixel or one 463 m MODIS pixel covers many cells. Thus, do not subtract the values of two adjacent cells and call the result a gradient.

What to do when a call fails

Recall does not accept the MCP task parameter. Only emem_eudr_dds and emem_hunt run as tasks.

One-call alternative

POST /v1/ask {"question":"what is the elevation of Bengaluru?"} (MCP emem_ask) classifies the question and runs locate and recall for you. It returns answer, fact_cids, facts_summary and a receipt. It does not return a facts[] list or a memory_token. Use the two-step path when you must control the bands or the place resolution.

Before you quote the number

Give the memory_token, not only the number. The receiver then resolves the byte-identical signed fact and does not have to trust your summary. See `emem-referential-drift`. Before you send the answer, run `emem-verify-before-publish`.