emem — Earth memory protocol

Resolves a place mention (free-text name, address, or lat/lng) to the protocol's cell64 identifier, and returns the topic-grouped inventory of bands and algorithms available at that location. When to use: Use whenever the input refers to a real-world location and the next step needs the cell64 identifier or wants to know which bands are available before recalling. The response carries `data_at_this_cell` with three sub-fields: `live_bands_by_topic` (every band recallable here, grouped by topic such as flood_water_event_window, vegetation_condition, built_up_human_geography), `algorithms_for_topic` (composition recipes that fuse those bands into named scores), and `declared_but_no_materializer_at_this_responder` (cube slots reserved without a live connector). For the single-shot path that runs the full chain server-side and returns one packaged answer, use `emem_ask` instead.

Single-shot free-text answer about a real-world location, backed by signed satellite/elevation/water/built-up receipts. Forwards a place mention plus a question; runs the locate → recall → algorithm chain server-side; returns one packaged envelope. When to use: Use when the question concerns a specific real-world place and a packaged, citation-bearing answer is preferable to manual primitive composition. Forward the user's question verbatim as `q` plus the location as `place` (free text), `cell` (cell64), or `lat`+`lng`. The server resolves the location, classifies the question to a topic, recalls every relevant band (auto-materializing Sentinel-2 / Sentinel-1 / Cop-DEM / JRC GSW / Overture / weather on miss), surfaces the algorithm recipes that compose those bands into named scores, and returns a single envelope with `topic_routing`, `facts`, `algorithms_for_question`, an optional Sentinel-2 RGB scene URL, and a `caveats` block (grid resolution, revisit cadence). All facts are signed by the responder; the signed `receipt` (and its content-addressed `fact_cids`) is surfaced at the envelope ROOT — `response.receipt` / `response.fact_cids` — exactly like every other primitive, and is also mirrored under `facts_summary.receipt` for back-compat. Set `include_image: true` to bundle the latest cloud-free Sentinel-2 thumbnail. Out-of-scope questions return `topic_routing.matched_topic: null` plus the full inventory so the caller can route elsewhere.

Event-discovery sweep: pick an event keyword (algal_bloom, deforestation, flood_extent, wildfire, urban_heat_island, methane_plume, landslide, drought, soil_salinity, crop_stress, water_turbidity, oil_slick) plus a region (free-text name or polygon_bbox). The responder geocodes the region, fans out across up to 32 sampled cells, recalls each event's primary scalar input band, and returns the top 8 hotspots ranked by that scalar — each carrying its cell64, lat/lng, the recalled value, a fact_cid for citation, and a scene.png URL. Bypass for free-text input is `emem_ask` (the classifier in /v1/ask routes "find X in Y" questions to the same hunter path). When to use: Call when the user asks an open-world discovery question ("find oil spills in the Persian Gulf", "where is deforestation happening in the Amazon", "show me algal blooms in Lake Erie", "hunt wildfires across California"). Surface 3–8 hotspots with their scene.png as image attachments and quote at least one fact_cid. For `oil_slick` the responder honestly reports `not_yet_implemented` and points at SAR-darkening + turbidity proxies — don't fabricate detections. The ranking uses the algorithm's primary scalar input only; for the full per-cell algorithm score, fetch the formula at /v1/algorithms/<key> and apply it client-side over the same recalled bands.

Produce a Due Diligence Statement per Regulation (EU) 2023/1115 for one or more plots. Each plot carries operator-supplied geometry (GeoJSON Polygon for >4 ha, Point for ≤4 ha non-cattle per Article 2(28)), country of production (ISO3), Combined Nomenclature code (HS-6+), and quantity in kg. The endpoint applies the regulation's 10 % canopy / 0.5 ha / 5 m height forest definition (Article 2(4)) using the EU Commission's expected JRC GFC2020 V3 baseline plus Hansen GFC v1.12 loss-year confirmation; Sims et al. 2025 driver attribution and RADD SAR fallback layer on when those connectors are wired (Absence today). The response is an Annex II-shaped envelope with per-plot verdict (pass/fail/not_in_scope/indeterminate/below_mmu), failing-cell fraction, and signed fact CIDs for every per-cell verdict — operators quote them in the company's Article 12 record. Article 9(1)(b) legality (land tenure, FPIC, country-of-origin laws) is structurally out of EO scope; the response carries an explicit `legality_disclaimer` for that reason. When to use: Call when a commodity supplier or EU importer needs to evidence due diligence under Regulation (EU) 2023/1115. Use the plot-level signed receipts as evidence inside the operator's company record; pair with a partner legality module before submitting the final DDS to the EU Information System (TRACES NT). For a single plot, pass one entry in `plots`. For batch supply-chain audits, pass up to a few dozen plots in one call — the endpoint fans out per plot. Surface the failing-cell fraction, the chosen forest baseline, and the legality disclaimer in the user-facing response so the operator understands what the engine claims (and does not).

Compute the Standardized Precipitation Index (McKee et al. 1993) at a cell: fit a gamma distribution to the same-window precipitation-accumulation history, then standardize the current accumulation to a z-score and map it to a drought class (extreme/severe/moderate drought … normal … wet). Supply `precip_history_mm` + `current_accumulation_mm` directly, or omit them to read the stored `weather.precipitation_mm` trajectory and build the window accumulations server-side. `window_days` selects SPI-1 (30 d), SPI-3 (90 d, default), SPI-12 (360 d), etc. When to use: Call when the user asks 'is this place in drought', 'how dry is it relative to normal', or wants a precipitation-anomaly z-score. The response is honest: when fewer than the WMO-recommended minimum samples exist it returns verdict=`inconclusive` with `spi:null` and a `honest_note` rather than fabricating a z-score from a handful of points. Quote the `spi`, `spi_class`, and `n_samples`. For raw precipitation use `emem_weather`; SPI is the standardized anomaly.

Compute the differenced Normalized Burn Ratio (dNBR = NBR_pre − NBR_post; Key & Benson 2006) and map it to the USGS burn-severity classes (unburned / low / moderate-low / moderate-high / high). Supply `nbr_pre` + `nbr_post` (pin the scenes bracketing the fire date) for a correct result, or omit both to use the two most-recent stored `indices.nbr` scenes (older=pre, newer=post) as a coarse estimate. When to use: Call after a wildfire to quantify how badly an area burned, or to triage post-fire severity across a region cell-by-cell. Best practice: explicitly pass `nbr_pre`/`nbr_post` from scenes that bracket the known fire date — the stored-trajectory fallback just takes the two most-recent scenes and may not bracket the fire. Surface `dnbr` and `severity_class`. For active-fire detection use `emem_hunt` with the wildfire event instead.

Estimate seasonal CH4 emissions from rice cultivation per IPCC 2019 Refinement Eq 5.1: integrate the daily emission factor over the cultivation period with water-regime scaling (SFp pre-season, SFo organic amendment) and an optional Yan-2005 Q10 temperature modifier. `cultivation_period_days` and the regional `efc_kg_ch4_ha_day` (Table 5.11) are REQUIRED — the endpoint refuses to guess a global default because the regional EFc drives the magnitude (~30% bias if wrong). An NDWI series (supplied or read from stored `indices.ndwi`) informs the flooding-regime context. When to use: Call for paddy-rice GHG inventory / MRV work where the user needs kg CH4 per hectare for a cultivation season. The caller MUST pick the IPCC region's EFc row (Table 5.11) and the cultivation-period length; pass SFp/SFo when the water regime or organic amendment is known. Surface the seasonal emission, the EFc used, and the scaling factors so the inventory is auditable. For enteric/fertilizer pathways use the dedicated sustainability endpoints.

Composite deforestation-alert score: `alert_score = 0.5·clamp01(ndvi_drop/0.30) + 0.5·clamp01(embedding_change/0.20)`, where `ndvi_drop = max(0, ndvi_modis_baseline − ndvi_now)` and `embedding_change = 1 − cos(tessera_latest, tessera_prev)`. Each half degrades INDEPENDENTLY and honestly: if a band is missing, that half is dropped AND the output is renamed so a half-score can never be mistaken for the full composite. If NEITHER half is computable the response is a signed `inconclusive` carrying no number. When to use: Call to flag recent forest-loss-like change at a known cell when you want a single 0..1 alert score rather than a full ensemble. Read the renamed score field and the present/absent halves — don't treat a half-score as the full composite. For multi-cell open-world discovery use `emem_hunt` (deforestation event); for the three-encoder change ensemble use `emem_triple_consensus`; for regulatory EUDR evidence use `emem_eudr_dds`.

Cloud- and night-independent Sentinel-1 C-band confirmation of forest disturbance. Intact forest scatters VV strongly + stably (canopy volume scattering); clearing collapses that term so VV backscatter DROPS ~3-5 dB. Samples VV at a baseline-year July-1 anchor and the latest scene, reports `vv_drop_db = baseline − recent` and a `disturbed` flag when the drop ≥ 3 dB (Reiche et al. 2018, RSE 204:147). Both VV reads are signed Primary facts; the response cites both fact_cids. Honest `inconclusive` when either S1 vintage is unavailable. Source: Microsoft Planetary Computer sentinel-1-rtc (anonymous SAS — no requester-pays, no API key). When to use: Call to corroborate or scout forest clearing where cloud blocks the optical products — radar sees through cloud and at night, catching wet-season clearing the annual Hansen/JRC-TMF layers and a single cloudy Sentinel-2 pass miss (the gap RADD was meant to fill). This is an ADDITIVE scout signal, NOT a standalone legal verdict: a VV drop can also be transient (soil moisture, harvest, flood recession), so confirm with the optical consensus (`emem_eudr_dds` or `emem_deforestation_alert`) before crediting a decision.

Three-encoder change ensemble: compute the cosine change between the two most-recent DISTINCT vintages for each of the Clay, Prithvi, and Tessera embeddings at the cell, then vote each encoder's change against `consensus_threshold` (registry default 0.15). Returns each encoder's change magnitude, its vote, and the consensus verdict (how many of the three agree change happened). Degrades to a signed `inconclusive` when the GPU sidecar is unreachable or a cell lacks two distinct vintages for the encoders. When to use: Call when the user wants a robust, model-agnostic 'did this place change' answer backed by three independent foundation encoders rather than one — e.g. cross-checking a single-encoder alert, or auditing change with consensus voting. Surface the per-encoder change + the vote count. When only one encoder has two vintages the verdict is honest about the thin evidence. For a single-encoder vector delta use `emem_state_diff`; for the NDVI+embedding proxy use `emem_deforestation_alert`.

Compute three standard DEM terrain indices from one 3×3 Copernicus-DEM (copdem30m.elevation_mean) neighbourhood at a cell: Horn (1981) slope in degrees, Riley (1999) Terrain Ruggedness Index (TRI = sqrt(Σ(Z_centre−Z_i)²)), and Weiss (2001) Topographic Position Index (TPI = Z_centre − mean(neighbours); positive = ridge, negative = valley). The 8 neighbour cell64s are derived by perturbing the cell's lat/lng one cell pitch per axis; the east-west ground spacing is cos(lat)-corrected. When to use: Call when the user asks how steep / how rugged / ridge-or-valley a place is, for siting (solar, construction, agriculture), erosion/landslide screening, or habitat-heterogeneity inputs. Slope and TRI need the full 8-neighbour ring; TPI degrades to ≥1 neighbour. Copernicus DEM is bathymetry-free, so ocean cells return a signed `inconclusive` rather than a fabricated slope — read each index's own `verdict`. For raw elevation use `emem_elevation`.

Answer 'how alike are these two places?' Mean-pool the 128-D GeoTessera embedding across each region's cells to get a centroid, then return the cosine similarity in [-1,1] (+1 = identical landscape, 0 = unrelated). Each region is {place} | {polygon_bbox} | {cells}. CPU-fetched embeddings — no GPU sidecar needed. Surfaces how many cells in each region actually carried a vector (coverage). When to use: Call to compare two areas at the level of overall land character (e.g. 'is this valley like that one?', 'find me somewhere that looks like X'). Degrades to a signed `inconclusive` (no number) when a region has no embedding-covered cells. For a single cell-to-cell vector cosine use `emem_compare`; for k-NN retrieval use `emem_find_similar`.

Mean-pool the 128-D GeoTessera embedding over a region's cells: centroid = (1/N) Σ v_i, plus the L2-normalised centroid and a content-addressed centroid_cid. The building block region_similarity composes. Region is {place} | {polygon_bbox} | {cells}. NaN dims are averaged over their finite contributors. CPU-only. When to use: Call when you need one representative embedding vector for an area — to feed similarity search, clustering, or a linear probe over places rather than single cells. Returns a stable centroid_cid for citation. Signed `inconclusive` when no cell in the region carried a vector.

Quantify how varied a region's landscape is: diversity = (1/(N(N-1))) Σ_{i<j} (1 − cosine(v_i, v_j)), the mean pairwise cosine distance over the region's GeoTessera embeddings. 0 = perfectly uniform; higher = more heterogeneous land cover (a determinantal-point-process / k-medoid diversity). Region is {place} | {polygon_bbox} | {cells}. CPU-only. When to use: Call for habitat-heterogeneity / biodiversity-proxy inputs, or to tell a monoculture from a mosaic landscape, or to rank regions by how mixed they are. Needs ≥2 embedding-covered cells, else a signed `inconclusive`. Pair with `emem_terrain` ruggedness for a fuller heterogeneity picture.

Score how much a cell looks like its surroundings: consistency = (1/8) Σ cosine(centre, neighbour_i) over the 8 immediate cell64 neighbours, plus outlier_score = 1 − consistency. High consistency = the cell blends in (Tobler's First Law); high outlier_score = it stands out — an edge, a fresh clearing, a built patch in farmland. CPU-only GeoTessera embeddings. When to use: Call to flag a cell that is anomalous versus its local neighbourhood (change/edge detection, QA of a homogeneous expectation, scouting for the odd-one-out). Signed `inconclusive` when neither the centre nor any neighbour carried an embedding. For year-over-year change at one cell use `emem_state_diff` or `emem_triple_consensus`.

Get one dense numeric fingerprint that summarises everything known about a place — ready to feed into similarity search, a classifier, or clustering. Two views: `encoder` returns a single AI-model embedding (128-D Tessera, 1024-D Clay, 1024-D Prithvi); `cube` returns the full 1792-D vector concatenated across every band, with a per-band coverage manifest. When to use: Call this when the user wants a machine-usable summary of a place rather than individual band readings — e.g. 'give me a feature vector for this location', 'how do I represent this place for ML', or before running similarity / linear-probe / clustering downstream. Also use it to get one rebindable handle (`memory_token` / `state_cid`) that cites the whole place. Default `view=encoder` is the cheap single-recall path; pass `view=cube` for the full attested view (its `coverage[]` lets you tell signed-zero from not-yet-materialised). Then hand the vector to `emem_find_similar` (k-NN), `emem_compare` (two-place cosine), or `emem_verify_receipt` (audit the signature).

Get the place's fingerprint from several AI models at once (`geotessera`, `clay_v1`, `prithvi_eo2`, `galileo`) in one call, returned as a per-model map. Each model is tried independently; any that can't produce a vector here show up under `missing` with a reason instead of failing the whole request. When to use: Call this when the user wants a second (or third) opinion on what a place looks like — 'do the different models agree this is forest / urban / water?', 'which model has the freshest read here?', or when you want all the embeddings concatenated for a stronger downstream classifier. Use the single-model `emem_state` instead when one embedding is enough. Pass `encoders: [...]` to narrow the set.

Vector delta between the same cell at two tslots: returns the per-element residual, its L2 norm (scalar change-magnitude), the cosine between the two source vectors (orientation drift), and both source fact CIDs so the agent can quote both attestations as evidence. When to use: Call when the user asks 'how much did X change between A and B' for a foundation embedding at one place. Pass `tslot_a` and `tslot_b` (must differ); default `encoder=geotessera`. For per-band scalar change (NDVI delta, elevation delta) use `emem_diff` instead.

Compose a `memt:<cell64>:<fact_cid>` (or `memt:<cell64>:<state_cid>`) citation handle. Validates both components are non-empty and do not contain the outer separator `:`. When to use: Call when the agent wants a single rebindable string to cite a place + attested fact across messages, threads, or tools. The token is the recommended way for agents to pass earth-memory citations to other agents without re-fetching. Pair with `emem_verify_receipt` on the receiving end to verify the signed payload.

Parse a `memt:<cell64>:<fact_cid>` citation handle and return the signed fact body the cid binds. Saves the agent from string-splitting the token and chaining `GET /v1/facts/<cid>` manually. When to use: Call when an agent receives a memory_token from another agent (or out of a previous turn) and wants the underlying signed bytes. The response carries the parsed cell + fact_cid, the full fact body, and the stable `fact_url` an agent can hand to any other peer. 404 with a typed code if the responder doesn't hold the cid; try /v1/fetch with the cid then, or paste the token at a mirror.

Compose N (cell, band, tslot?) triples into ONE signed envelope. Each triple runs through the standard auto-materialize recall path; the resulting fact_cids are bundled into a content-addressed envelope and the responder signs over the full receipt. The composed `bundle_token` is `memb:<bundle_cid>` — a single rebindable string that cites the whole set. When to use: Call when the agent wants to cite multiple (place, band, vintage) facts as one handle. The bundle stays verifiable offline via /v1/verify_receipt (the receipt covers all cited fact_cids and cells). Use this instead of N separate `emem_memory_token` composers when the citation is conceptually one thing (e.g. "the EUDR-relevant baseline for these 8 plots at 2020-12-31").

Parse a `memb:<bundle_cid>` token and return the signed bundle envelope: every citation (cell, band, resolved_tslot, fact_cid, memory_token), the receipt, the responder pubkey, and the deduped flat cells[] / fact_cids[] arrays. Returns 404 with a typed code when the responder does not hold the bundle. When to use: Call when an agent receives a `memb:` token from another agent (or earlier turn) and wants the underlying signed citation set. The response is byte-identical to what `emem_memory_bundle` returned at the original responder.

Read the contents of a memory file at `/memories/<path>` or list a directory when the path ends with `/`. Optional `view_range: [start, end]` slices a 1-indexed inclusive line range out of the file. Mirrors the `view` verb in Anthropic's context-management-2025-06-27 memory tool spec. When to use: Call when the model running with `betas: ['context-management-2025-06-27']` issues a `view` against its memory directory. Use `/memories/` (trailing slash) to enumerate files; `/memories/notes.md` to read one. Returns a 404 with typed code on missing path.

Write a memory file at `/memories/<path>` with the supplied `file_text`. Overwrites if the file exists. Persists to sled, content-addresses the bytes (`file_cid`), and signs the write so the operation carries a verifiable receipt. Mirrors the `create` verb in Anthropic's context-management-2025-06-27 memory tool spec. When to use: Call when the LLM issues a `create` against its memory directory (initial scratchpad write, refresh of a notes file, etc.). The response carries the new `file_cid` and a signed receipt the agent can quote in audits.

Replace `old_str` with `new_str` in the named memory file. Fails (no partial write) when `old_str` is absent or matches more than once. Writes a new content-addressed `file_cid` and signs the receipt. Mirrors the `str_replace` verb in Anthropic's context-management-2025-06-27 memory tool spec. When to use: Call when the LLM issues a `str_replace` against its memory file — typical for small targeted edits. The strict single-match contract is the contract Claude expects: an LLM that sees a single-match diff knows the change applied where it intended.

Insert `new_str` after the given 1-indexed line in the named memory file. `insert_line: 0` inserts at the top. Writes a new `file_cid` and signs the receipt. Mirrors the `insert` verb in Anthropic's context-management-2025-06-27 memory tool spec. When to use: Call when the LLM wants to append a new line to a memory file without rewriting it. For top-of-file inserts, pass `insert_line: 0`; for end-of-file, pass the current line count (the responder rejects out-of-range with a typed error).

Delete a memory file at `/memories/<path>`. When the path ends with `/`, every file beneath the directory is removed. Updates the path index but leaves prior content-addressed blobs in place (the audit history is append-only). Mirrors the `delete` verb in Anthropic's context-management-2025-06-27 memory tool spec. When to use: Call when the LLM issues a `delete` against a memory file or subdirectory it no longer needs. Existing receipts citing the old file_cid stay verifiable — the blob is content-addressed, only the path → file_cid index forgets.

Move (rename) a memory file from `old_path` to `new_path`. Both paths must stay under `/memories/`; `new_path` must not already exist. The file_cid is preserved (no re-sign) so the prior receipt still binds the bytes. Mirrors the `rename` verb in Anthropic's context-management-2025-06-27 memory tool spec. When to use: Call when the LLM wants to rename or move a memory file. Failure modes: source missing, destination already exists, path escapes `/memories/`.

List memory files by their typed `kind` (episodic | semantic | procedural | resource). Optional path prefix narrows the scan; results are sorted by signed_at descending. The kind taxonomy follows the CoALA / LangMem / MIRIX agent-memory ontology: `episodic` = observations of events, `semantic` = durable learned facts, `procedural` = playbooks, `resource` = generic durable scratchpad (default for back-compat). When to use: Call when an agent wants only one slice of its memory (e.g. surface every semantic fact it has learned about a topic) without scanning the full directory tree. Pair with memory_view for read-back of a specific entry.

Semantic search over /memories/* file contents using BGE-base-en-v1.5 (768-D, L2-normalised) backed by a Lance partition (`memory_text_index_d768.lance`). Matches paraphrases — "rainfall in March" finds "precipitation observed in spring" without an exact substring match. Returns ranked hits with similarity in [0,1], 200-char snippets around the best-matching chunk, and the signing receipt's path / file_cid / signed_at / attester_pubkey_b32 fields. Filters: `kind`, `path_prefix`, `attester_pubkey_b32`. Falls back to a brute-force scan (slower but correct) when the index is empty or `EMEM_DISABLE_LANCE=1` is set; the `via` field of the response reports which path was taken. When to use: Call instead of paging through `memory_view` whenever the agent knows roughly what it wants (a topic, a name, a paraphrase) but not the exact file path. Pair with `memory_view` for the full body once you've narrowed down the candidate — `emem_memory_search` returns a 200-char snippet, not the whole file. The polling indexer hydrates once per minute (configurable via `EMEM_MEMORY_SEARCH_POLL_SECS`), so a file created in the same turn may briefly miss the fast-path — the brute-force fallback still catches it.

Recall facts about a cell — auto-materializes on miss for any band with a registered materializer. When to use: Call after `emem_locate` (or with a known cell64). Returns every Primary fact stored at that (cell, band, tslot). IMPORTANT: if the cell has no fact yet for a requested band AND that band has `has_materializer=true` (per `emem_coverage_matrix` / `emem_materializers`), the responder fetches the upstream value, signs it under its identity, persists it, and returns it in the same response (~180 ms first call, ~10 ms cached thereafter). So for any wired band you can recall ANY cell on Earth without seeding — just pass `bands: [<band>]`. The response carries `materialize_notes` listing what was just fetched. Empty result with no notes means the band has no materializer at this responder.

Recall facts across every cell inside a place's polygon (single signed envelope). Closes the place-name-drift gap for wide features (parks, lakes, regions). When to use: Call when the user names a wide feature (national park, river basin, country, large urban area) where one cell is too small. Pass `place` and the geocoder will fan out across the polygon — or pass `polygon_bbox` directly if you have coordinates. Returns `merged_facts`, `by_cell`, and a `polygon_bbox.source` indicator (`nominatim_boundingbox` = real polygon, `centre_cell_bbox` = fallback to one cell because the geocoder had no polygon). For *farm* queries the OSM polygon is the whole estate envelope; pass `include: ["ftw_fields"]` to additionally attach per-field agricultural-boundary polygons from Fields of The World (CC-BY-4.0) — or call the dedicated `emem_field_boundaries` for the pure-fetch shape.

Per-field agricultural-boundary polygons from the Fields of The World global product (~3.17B fields, 241 countries, 10 m resolution, CC-BY-4.0). Returns a GeoJSON FeatureCollection with the polygon geometries, FIBOA-compatible properties, and a planar `area_m2` per field — plus provenance (source CID, provider URL, license, attribution). When to use: Call when the user asks about farms, fields, parcels, croplands, plots, or agricultural boundaries inside a region — anywhere the OSM/Nominatim boundary alone is too coarse (the OSM polygon for a farm is its estate envelope; this returns the individual field polygons inside). Pass `place` (free-text) or `polygon_bbox`. For farms wider than ~10 km², split the bbox: the fetcher caps each call at 16 covering tiles. The receipt quotes `license: CC-BY-4.0` and `attribution: Fields of The World / Taylor Geospatial Institute` — surface both with any rendered map. For a one-shot "facts at every cell inside the farm PLUS the field polygons", call `emem_recall_polygon` with `include: ["ftw_fields"]` instead.

Query facts over a region (single cell or list of cells), optionally aggregated per band. When to use: Call when the user asks 'how does region X look', 'what's the average NDVI here', or wants a region-level summary. Use `agg=mean|median|p90|vector_centroid` to fold per-band values.

Compare two cells: cosine similarity over shared vector bands + per-band scalar deltas. When to use: Call when the user asks 'how similar is X to Y', 'compare these two places', or wants a difference vector. Returns a single cosine score and per-band deltas.

Compare two bands at the same cell. Scalar pair → metric=delta, value=b-a. Vector pair (equal dim) → metric=cosine + per-dim delta. Returns a signed receipt naming both source fact CIDs. When to use: Call when the user wants cross-source consistency at one place ('does Cop-DEM agree with GMRT here?'), cross-vintage drift ('how did the embedding change between 2017 and 2024 at this cell?'), or any band-vs-band comparison within a single cell. `cell` + `a` + `b` are required. `tslot_a`/`tslot_b` are OPTIONAL: omit them to let the responder auto-pick each band's latest attested tslot — required for medium/fast-tempo bands (NDVI 30-day, MODIS 8-day, weather, CAMS) where there is no fact at tslot=0. The response carries `tslot_resolution` (echoes what was chosen and why) and `bands_with_no_history` (lists any band the cell has no attested fact for).

k-NN over the corpus by cell embedding or inline vector. When to use: Call when the user asks 'find places like X', 'where else looks like this', or hands an embedding to find neighbours. `key` is either a cell64 or `inline:[x,y,...]`. Default band is `geotessera` (128-D Tessera foundation embedding); pass `band: "geotessera.multi_year"` for the 1152-D 9-vintage (2017–2025) fusion.

Time series for one (cell, band) over an inclusive [start, end] tslot window. Returns only what's already attested — does NOT trigger materialization. For historical backfill use `emem_backfill`. When to use: Call when the user asks 'how did X change over time' for a band that already has multiple historical tslots seeded. IMPORTANT differences from `emem_recall`: (1) trajectory does NOT auto-materialize past tslots — it returns only facts that have already been attested at this responder, so for fast-tempo bands like `indices.ndwi` you'll typically see ONE point at the latest tslot until an attester seeds history. (2) tslots are non-negative `u64`; there's no negative-offset 'last 2 years' shorthand. For LONG-TERM history questions ('flooded in last 2 years', 'forest loss since 2020') prefer either (a) a static-tempo summary band that one fact answers — `surface_water.recurrence` covers 1984-2021 in a single signed value, no trajectory needed — or (b) `emem_backfill` to materialize and sign the missing tslots in one call.

Compute a DerivativeFact (delta) between a band's values at two tslots. When to use: Call when the user asks 'what changed between t1 and t2', 'give me the delta'. Returns a signed DerivativeFact + receipt — the delta itself is content-addressed and citable.

Surface where the corpus DISAGREES with itself. When two or more independent sources signed different values for the same place + band + time, this returns that disagreement with a 0–1 severity score and citations to every disputed fact — instead of silently picking one value and hiding the conflict. The opposite of a confident single answer: it tells you when not to trust one. When to use: Call this when trust matters before you rely on a number — 'is there disagreement about X', 'do the sources corroborate this', 'audit this claim', or 'find contradictory observations in region Y'. Use it to decide whether a fact is well-corroborated or contested. Narrow with `cell_prefix` (e.g. "defi.zb5") for a region and `band` for one family; `min_severity` filters out trivial differences. Severity is per band kind: scalar = spread over the band's range, vector = 1 − mean cosine, categorical = 1 − mode share. The receipt cites every disputed CID — follow up with `emem_diff` to quantify a pair, or (with the refinement loop on) read the emitted `disagrees_with` edge via `emem_edges_recall`.

Read temporal knowledge-graph edges (subj --pred--> obj, valid over [valid_from, valid_to)), bi-temporally filtered, in EITHER direction. Forward (`subj`, direction="out", the default): edges originating at a subject fact. Reverse (`obj`, direction="in"): edges pointing AT a fact — what disagrees-with / supersedes / relates-to it. Returns a signed list of edges plus the distinct neighbour fact CIDs (`objs` for out, `subjs` for in); the receipt commits the returned edge CIDs into its signature preimage. When to use: Call this to read the typed CONNECTIONS of a fact — what disagrees with it, what superseded it, what relates to it — as of a point in time. A plain recall gives you the fact; this gives you how that fact links to others in the memory graph. Ask it when the user says 'what is this related to', 'what replaced this observation', 'why is this value contested', or 'what did this place's relations look like as of date X'. Pick a direction: set `subj` (direction="out") to ask 'what does this fact point at'; set `obj` (direction="in") to ask the REVERSE — 'what disagrees-with / supersedes / points-at this fact'. Set exactly one of subj/obj — an ambiguous or empty request errors honestly rather than returning a silent empty. Pass `as_of_tslot` to get the latest edge per neighbour whose valid interval covers that moment (newer edges shadow older — nothing is deleted); pass `pred` (e.g. `disagrees_with`, `supersedes`) to filter, or omit it (empty string) for every predicate. Tip: a quicker way to get a fact + its outbound edges in one shot is `emem_recall` with include:["edges"]. Follow each edge's `obj`/`subj` with `emem_fetch` to resolve the related fact, or `emem_verify_receipt` to confirm the signature offline.

Fetch a fact by its content-address (CID). Returns the full signed Primary or Absence fact — the same body served by REST `/v1/facts/{cid}`. Closes the citation loop: any fact_cid surfaced by recall, materialize, attest, or verify can be re-resolved by another agent without REST. When to use: Call whenever you have a `fact_cid` (e.g. from `emem_recall`'s response, an `emem_attest` receipt, an `emem_materializers` outcome, or a citation in another agent's reply) and need the full fact body — its value, unit, sources, signer, signed_at, and derivation. Particularly useful for verifying that a citation a downstream agent gave you actually resolves on this responder. The response is byte-identical across responders for the same CID — the CID itself is the validator.

Materialize and sign every per-tslot fact for one (cell, band) inside a [start_unix, end_unix] window. Returns a signed list of (tslot, fact_cid, status) for each step. Slow but possible — one upstream fetch per tslot, capped by `max_facts`. When to use: Call when the user wants HISTORY for a fast/medium-tempo band and `emem_trajectory` returned only the latest point. The responder iterates the tslot range derived from the band's tempo, calls the per-tslot historical materializer, signs each result, and persists. After completion `emem_trajectory` over the same window returns the full series. Bands without a historical materializer (e.g. `weather.*` from met.no's nowcast) return `status: "present_only"` for past tslots — check `emem_coverage_matrix.history_available_from`/`history_available_to` to see how far back each band can be backfilled. Prefer this over staking an attestation when the upstream is publicly fetchable.

Forward-step 2-D explicit finite-difference solver for the heat equation ∂u/∂t = α∇²u over a 3×3 cell stencil centred on `cell`. Reads `modis.lst_day_8day` (Land Surface Temperature) at the centre and 8 cell64 neighbours, integrates N hours ahead under a CFL-stable timestep, returns a signed forecast. Real PDE rollout — not a decay-scoring heuristic. When to use: Use when the user wants a short-horizon LST forecast (urban heat island, surface-temperature evolution, heatwave onset modelling) at a specific cell. Default α=1e-6 m²/s matches urban surface diffusivity (Oke 2017); pass a smaller α for water bodies or higher for vegetated surfaces. The solver caps at one-week horizons because the 8-day MODIS composite stops being a representative initial condition past that. Each call materialises 9 MODIS facts (one per neighbour) on miss — first call ~5 s cold, ~30 ms warm. Receipt cites all 9 input fact CIDs.

Forward-step 1-D explicit finite-difference solver for the shallow-water wave equation ∂²u/∂t² = c²∂²u/∂x² with c² = g·h, where depth h comes from `gmrt.topobathy_mean` along the seaward gradient. Models how an offshore swell of height H_s and period T propagates toward `coastal_cell`. Returns arrival height + time + depth + phase-speed profiles, all under a CFL-stable timestep. When to use: Use when the user wants to predict swell arrival at a coast (storm-surge planning, shoreline-impact assessment, surf forecasting). The solver walks `n_offshore_cells` cells seaward from `coastal_cell` along the bathymetric gradient (default 8 cells = 80 m of profile at the active 10 m grid), samples GMRT depth at each, and integrates the wave equation forward until the wavefront reaches the coast plus one period. Receipt cites every depth fact CID along the profile. Returns 422 with a clear message if `coastal_cell` is land-locked.

Predict next-month NDVI at a cell using a constrained JEPA-pattern AR(2) seasonal predictor. Reads up to 24 past months of `indices.ndvi`, fits a closed-form predictor `y_{t+1} = α·(lag-12 NDVI or recent mean) + β·(last + slope) + γ·recent_mean`, returns the prediction clamped to NDVI's physical range. Coefficients (α=0.6, β=0.3, γ=0.1) are NOT learned — they're fixed from the agricultural-NDVI literature. v2 (future) will train an actual encoder + predictor on the geotessera embedding pool. When to use: Use when the user wants a one-month-ahead NDVI forecast at a specific cell (crop-stress monitoring, growing-season tracking, vegetation-anomaly anticipation). Lookback defaults to 6 months; if fewer monthly tslots are attested at this cell, the predictor uses what's there and surfaces the count in `lookback_months_used`. Returns 422 if no NDVI history exists at the cell — chain to `emem_backfill` first to seed history. Receipt cites every input NDVI fact CID.

Predict the next-step value of 4 environmental scalars at a cell — `indices.ndvi`, `modis.lst_day_8day`, `modis.lst_night_8day`, `cams.pm25` — using a small learned dynamics MLP. Reads up to K=6 most-recent attested lags per band, runs them through an ONNX dynamics head (~200k params, CPU-fast), and returns a per-band {value, confidence, n_real_lags, via}. The receipt's `model` block carries `model_id`, `version`, `blake2b_hex` (model_cid), training/validation provenance, a top-level `skill_vs_persistence` block, and `honesty_warnings` — flagging `untrained_baseline` when the artifact is the zero-init sentinel and `NEGATIVE_SKILL` when the learned model is worse than persistence on real held-out NDVI. When the model does not beat persistence, bands with a real lag are returned from that lag tagged `via:persistence_fallback_negative_skill` (bands with no real lag fall back to labelled climatology). Distinct from v1 (`emem_jepa_predict`) which returns a single NDVI scalar via closed-form coefficients. When to use: Use when you want a short-horizon forecast of NDVI / land-surface temperature / PM2.5 at a cell grounded in its attested history. Returns 422 with a `/v1/backfill` hint when the cell lacks enough cached lags. Always read the receipt's `model.honesty_warnings` — `untrained_baseline` means the trivial 'predict last vintage' baseline (treat as no-op), and `NEGATIVE_SKILL` means the served values are the persistence fallback, not a learned improvement. Check each band's `via` field to see whether its value came from the learned model, persistence, or climatology.

Verify a structured claim against a cell's facts. Returns verdict + evidence CIDs + signed receipt. When to use: Call when the user asks a yes/no question about a cell ('is the NDVI > 0.7 here', 'has this been deforested'), or when downstream code wants citable evidence for a logical predicate.

Per-key drill-down on a single composition recipe — full body (kind, inputs, formula, output, citation, references) for ONE algorithm key. Companion to `emem_algorithms` (which is the catalog). When to use: Call when you already know the algorithm key (from `emem_algorithms`'s catalog or the topic registry) and need its full math. Cheaper than fetching the 190 KB catalog when you only need one entry. Returns the same structure that `/v1/algorithms/{key}` does. 404s with `cid_not_found` if the key isn't registered — call `emem_algorithms` for the live key list.

True-colour Sentinel-2 L2A RGB thumbnail centred on a cell. PNG returned as a native MCP ImageContent block (mimeType image/png). Pure-Rust pipeline: STAC search + HTTP-Range COG reads + 2-98 percentile stretch + PNG encode. When to use: Call when the user wants a VISUAL of a place — 'show me what this looks like', 'before/after the flood', 'is there a forest here', 'is this developed'. Returns a 256×256 px RGB image (~2.56 km × ~2.56 km at S2's 10 m native resolution), centred on the cell. Pass `cell` as a cell64 string OR a place name (auto-resolved). `max_cloud` filters scenes by `eo:cloud_cover` (default 20 %); raise it (60–80 %) for cloud-prone tropics if you keep getting 'no scene' errors. `datetime` is an RFC 3339 interval like `"2024-01-01T00:00:00Z/2024-12-31T00:00:00Z"` for a temporal slice (defaults to last 90 days). `structuredContent` carries the STAC item id, capture time, cloud_cover, EPSG, and per-channel reflectance percentile stretch values used — quote those alongside the image so the receipt is reproducible.

Cell polygon as a native MCP EmbeddedResource (mimeType application/geo+json). Properties carry centre lat/lng, bbox, approx size in metres, and the 8-cell neighbourhood — drop straight into Mapbox / Leaflet / Deck.gl / QGIS without a GIS pipeline. When to use: Call when the agent (or a downstream renderer) needs the cell as geographic geometry — for map overlays, polygon-clipping ops, or feeding a styling pipeline. Pass `cell` as cell64 or place name. The result is a GeoJSON Feature with Polygon geometry; for a FeatureCollection that includes every recalled fact's value as a property, fetch /v1/cells/{cell64}/recall_geojson?bands=... over plain REST instead.

Recall facts across a list of up to 256 cell64 strings in one signed envelope. Server fans out per-cell recalls in parallel, then aggregates the response. Auto-materializes any cell with a missing fact whose band has a registered materializer — same contract as emem_recall. When to use: Use after emem_find_similar (give it the neighbour cells), after emem_recall_polygon (when you want a deterministic cell list rather than a polygon), or whenever you have a precomputed set of cells (e.g. an admin-2 sample frame) and want one round-trip. Pass `cells: [c1, c2, ...]` plus the same `bands` shape as emem_recall. For more than 256 cells, batch the call.

One-shot elevation answer that fuses Cop-DEM 30 m (land), GMRT (ocean topobathy), and ESA WorldCover (water mask) into a single signed scalar at a place or coordinate. Returns `elevation_m`, the source actually used, and a `coherence_note` when the two surfaces disagree at the coast. When to use: Use when the user asks 'how high is X' or 'what's the elevation at this lat/lng' and you want the correct answer regardless of whether the cell is land, water, or coastline — the handler picks Cop-DEM for land and GMRT for water and surfaces the choice. Pass `place` (free text), `lat`+`lng`, OR `cell`. Otherwise, prefer emem_recall with `copdem30m.elevation_mean` / `gmrt.topobathy_mean` individually.

Turn a time-shaped question into a ready-to-run recall plan: it figures out WHICH bands to pull at WHICH past time windows (e.g. 'the year before the flood', 'last growing season', 'two vintages to compare') so you don't have to compute tslot offsets by hand. Returns the band + lookback + a `purpose` tag for each step. When to use: Call this first when the user's question is about CHANGE OVER TIME or a PAST EVENT and you're not sure which bands/dates to recall — 'was this flooded last year', 'what was the NDVI baseline before the fire', 'compare this place across vintages'. It hands you the recipe; then run those steps with `emem_recall`. Skip it when the user wants a single current reading. Pass `cell` plus an optional free-text `intent` hint. The plan is deterministic and the receipt cites which algorithm supplied each step.

Verify a signed receipt envelope server-side: recomputes the canonical preimage (`request_id | served_at | primitive | cells, | fact_cids,`), runs ed25519 over the embedded pubkey + signature, and returns `{valid, reason, pubkey_b32}`. Use when the in-browser /verify path is blocked (CDN offline, agent runtime has no crypto) or when you want a server-side audit of a third-party receipt. When to use: Pass a receipt object exactly as returned by any read primitive (signature can be byte[] or sig_b32; pubkey can be byte[] or responder_pubkey_b32 — the verifier tolerates both shapes). Optionally override `pubkey_b32` to assert verification against a specific signer. Returns 200 with `valid: false` when the signature fails — never 4xx for a structurally-well-formed bad signature.

One-shot multi-band recall at a place (or lat/lng). Defaults to emem's standard at-a-glance band set; pass `band` / `bands` to override. Polygon-resolved places stay at the centroid by default (`n_cells: 1`) to keep multi-band calls cheap — pass `n_cells: 2..=64` to fan out. When to use: Use when the user names a place and wants the standard situational readout (vegetation + elevation + landcover + recent weather) without picking bands. Polygon-aware: `place` that resolves to a polygon (park, lake, district) lands at the centroid unless `n_cells` widens it. For a single band, use the domain-specific shortcuts (emem_ndvi, emem_air, …) or emem_recall directly.

Recall Sentinel-2 NDVI (indices.ndvi, 10 m native) at a point or place. Composes locate → cell64 → recall in one call; auto-materializes on miss. When to use: Use when the user names a place (or lat/lng) and just wants the NDVI number. Polygon-resolved places default to a 16-cell fan-out aggregated as mean/median. Set `n_cells: 1` for point behaviour. For multi-band batches use emem_recall.

Recall Copernicus CAMS air-quality bands at a place: PM2.5 + NO2 + O3. Composes locate → recall → aggregate. When to use: Use when the user names a place and asks about air quality, pollution, or emissions exposure. CAMS is the European reanalysis — global coverage, ~0.4° native (resampled). For finer-grained urban PM2.5, pair with /v1/at-style stations data when available.

Recall MODIS land surface temperature day-8day + night-8day composites at a place. 1 km native, 8-day composite. When to use: Use when the user asks about surface heat, urban heat island, thermal anomalies, or wants day/night LST. Returns both fluxes so the agent can derive day–night spread.

Recall SoilGrids 250 m profile at a place: SOC, pH, clay/sand/silt fractions, bulk density, nitrogen — all at 0–30 cm depth. When to use: Use when the user asks about soil quality, agricultural suitability, or carbon stocks at a location. Six bands returned in one envelope.

Recall surface-water signals at a place: JRC Global Surface Water recurrence (1984–2021) + Sentinel-1 SAR backscatter (current). Pair detects standing water through clouds. When to use: Use when the user asks about flooding, wetlands, surface-water dynamics, or wants a robust water-presence check. JRC alone gives historical baseline; Sentinel-1 gives current flood detection.

Recall forest signals at a place: Hansen Global Forest Change (tree cover 2000 baseline + year-of-loss) + ESA WorldCover 2021 land class. When to use: Use when the user asks about deforestation, canopy cover, forest loss, or wants a forest-vs-not classification. Hansen gives year-of-loss for any cell with disturbance since 2001; WorldCover gives the current land class.

Recall the standard met.no/CAMS weather bundle at a place: 2 m temperature + total cloud cover + precipitation + 10 m wind speed. When to use: Use when the user names a place and asks 'what's the weather' or wants a now-cast snapshot. weather.* bands are now-only (no backfill); for climatology use terraclimate.*.

Submit a typed Intent; receive a plan or executed result. When to use: Call when the user asks something like 'where is X' or 'is A like B' and you don't want to pick a primitive yourself — the planner maps Intent variants to the right tool call.