The embedded multimodal database for embodied AI & robotics

Built for embodied-AI workloads

Not a trimmed general-purpose embedded database — every capability is shaped by the real anatomy of robotics data.

One SQL across every modality

Vectors, time series, full text, binary — not four databases, but four column types in one table. Real queries look like this:

-- 5 semantically nearest rows in the pose table (index-accelerated)
SELECT ts, sid FROM poses
ORDER BY emb <-> [0.1, 0.2, 0.3, 0.4]
LIMIT 5;

-- 60-second window aggregation + automatic TTL expiry
SELECT TIME_BUCKET('60s', ts) AS bucket,
       AVG(v), MAX(v)
FROM poses GROUP BY bucket;

ALTER TABLE poses SET TTL = '30d';

-- Hybrid retrieval: full text (BM25) × vector (KNN), RRF fusion
db.hybrid_search(
  "docs_body",            # full-text index
  "joint calibration",    # text query
  "docs_emb",             # vector column
  [0.1, 0.4, 0.2, 0.9], k=10)

-- BLOB byte-exact equality + one-shot MCAP import
SELECT id FROM frames
WHERE payload = X'DEADBEEF';

IMPORT MCAP 'session_2026_10_17.mcap';

-- Whole database → single file (CRC-verified + zstd)
EXPORT DB TO 'fleet.mote' COMPRESSED;

-- Deploy side: open read-only, multiple readers co-exist
db = Database.open_readonly("fleet.mote")

// All syntax is real — same as repository tests, not mockups

Designed for the AI stack — and readable by AI

The downstream of embodied intelligence is AI: semantic retrieval, training pipelines, agents. MoteDB builds these in rather than leaving them to the user.

# Semantic hybrid search: BM25 × KNN, RRF fusion
res = db.hybrid_search("docs_body", "joint calibration",
                       "docs_emb", [0.1,0.4,0.2,0.9], k=10)

# Results straight to Arrow → zero-copy into the training pipeline
tbl = db.query_arrow("SELECT ts, sid, emb FROM poses")
# tbl vector column is FixedSizeList<float32> — feed a DataLoader directly

5 minutes: from MCAP to a queryable time-series store

# pip install motedb-python
from motedb import Database

db = Database("robot.mote")
db.execute("CREATE TABLE poses (
  ts TIMESTAMP, sid INT, emb VECTOR(4)
) TIMESERIES(ts)")
db.executemany("INSERT INTO poses VALUES (?,?,?)", poses)

# Vector KNN: 5 nearest poses
hits = db.query("SELECT ts, sid FROM poses
  ORDER BY emb <-> [0.1,0.2,0.3,0.4] LIMIT 5")

# Transport: pack the whole database into one snapshot file
db.execute("EXPORT DB TO 'fleet.mote'")

Every line written for edge devices, validated to kernel standards

Edge failures cannot be rolled back or hot-fixed remotely. So releases pass through these gates — not marketing numbers, reproducible checks in the repository.

Documentation

Maintained in-repository, versioned with the code.