emem: A Content-Addressed, Verifiable Earth-Memory Protocol for AI Agents over Foundation-Model Embeddings

LLM agents asked “what is on the ground at this place, right now” have no stable handle for a patch of Earth and no way to prove the answer they return; the same agents asked “what did we learn here” fall back on unsigned, per-session scratchpads a server can rewrite. We present emem, a protocol giving both questions one trust surface. Every spatial fact is keyed by a 64- bit geographic cell (cell64, about 9.55 m at the equator), a band, and a temporal slot (tslot); the fact is serialized to deterministic CBOR, content-addressed by a truncated BLAKE3 digest (a 26-character base32 fact_cid), and returned inside an Ed25519 receipt over a domain-separated, length-prefixed preimage that any party can verify offline against the responder’s public key, with-out trusting the issuer. Facts are grounded in frozen Earth-observation foundation embeddings: Tessera (128-D, CPU-streamed from Cloud-Optimized GeoTIFFs), Clay v1.5 and Prithvi-EO-2.0 (1024-D, GPU-pinned), and Galileo (multimodal), whose independent receptive fields drive a triple-encoder change-consensus algorithm in which agreement is signal and a lone vote flags receptive-field aliasing. A read layer offers bi-temporal recall, Lance IVF_PQ k-NN with a TurboQuant binary-rotation Hamming fast path, region analytics, and a temporal-freshness kernel; a memory layer adds Anthropic-memory-tool file verbs over a CoALA-typed, capability-bound scratchpad, signed temporal edges, multi-attester contradiction scoring, and episodic-to-semantic merge with an opt-in sleep-time refiner. Cold cells materialize lazily on first request: any cell on Earth answers citeably with no pre-seeded corpus, and absence is itself a signed, typed, content-addressed receipt rather than a 404. We describe the protocol, its mathematics, a single-binary Rust reference implementation serving mirrored MCP and REST surfaces, and a mapping onto recent agent-memory literature.