Memory for AI,
owned by no one.

Engram is a decentralized vector database on Bittensor. Store embeddings with cryptographic proofs — no AWS, no central authority.

Open Dashboard Start Mining
◆recall@K scoring◆HMAC-SHA256 proofs◆Kademlia XOR routing◆HNSW indexing◆PyO3 Rust core◆TAO emissions◆content-addressed CIDs◆subnet netuid◆FAISS · Qdrant◆recall@K scoring◆HMAC-SHA256 proofs◆Kademlia XOR routing◆HNSW indexing◆PyO3 Rust core◆TAO emissions◆content-addressed CIDs◆subnet netuid◆FAISS · Qdrant

// how it works

Four steps.
Fully decentralized.

01SHA-256 · deterministic · content-addressed
ingest.py

Ingest

Send any text or pre-computed vector. Engram hashes it into a permanent CID — identical inputs always yield the same identifier, forever.

→Works with raw text (auto-embedded) or pre-computed float vectors
→CID format: v1::<sha256-hex> — 100% reproducible from content
→Optional metadata stored alongside the embedding
ingest.py
from engram.sdk.client import EngramClient
client = EngramClient(netuid=42)
cid = client.ingest(
    "Transformers changed NLP in 2017",
    metadata={"source": "arxiv", "year": 2017}
)
# → "v1::a3f9d2e8c7b14f09d6e3..."
# Same input always → same CID
assert client.ingest("Transformers changed NLP in 2017") == cid

// capabilities

Built different.

Everything you expect from a vector DB — plus cryptographic guarantees no centralized system can offer.

FeatureDescriptionStatus
Content-Addressed CIDsSHA-256 fingerprint per embedding — identical data always maps to identical CIDlive
HNSW IndexFAISS & Qdrant approximate nearest-neighbor — sub-50ms query at any scalelive
Kademlia DHT RoutingXOR-distance deterministic routing — same CID always routes to same minerslive
Storage ProofsHMAC-SHA256 challenge-response — validators slash miners who cannot prove storagelive
Rust Core (PyO3)CID generation + proof verification in compiled Rust — 10–50× faster than Pythonlive
TAO Incentivesscore = 0.50·recall@K + 0.30·latency + 0.20·proof_rate → TAO emissionslive
Replication ManagerAuto-detect degraded CIDs and trigger repair across redundant minersbeta
SDK / Python ClientEngramClient — drop-in for Pinecone, Weaviate, or any vector storelive
Scoring Formula — how miners earn TAO
50%
recall@K
Fraction of correct CIDs returned in top-K query results
30%
latency
Query response time — faster miners score proportionally higher
20%
proof_rate
Fraction of storage challenges answered with a valid HMAC proof

// developer SDK

Replace Pinecone
in an afternoon.

One Python client. Works with any embedding model. No API key, no vendor lock-in.

PineconeEngram
Open source✗✓
No API key✗✓
Storage proofs✗✓
Censorship-resistant✗✓
Self-hostable✗✓
TAO incentives✗✓
install
pip install engram-subnet
example.py
from engram.sdk.client import EngramClient
# Connect to the Engram subnet
client = EngramClient(netuid=42)
# Ingest text → returns content-addressed CID
cid = client.ingest(
    "Transformers revolutionized NLP in 2017",
    metadata={"source": "arxiv", "year": 2017}
)
# → "v1::a3f9d2e8c7b14f09..."
# Semantic search → top-K by cosine similarity
results = client.query("attention mechanisms", top_k=5)
for r in results:
    print(f"{r['score']:.4f}  {r['cid'][:24]}...")
# Vector search (bypass embedding step)
results = client.query_by_vector(my_vector, top_k=10)

// participate

Earn TAO.
Run the network.

Miners and validators earn from subnet emissions. Performance = yield.

Miner41% pool

Store embeddings, serve queries, pass storage proof challenges.

RAM4 GB minimum
Storage100 GB SSD
RuntimePython 3.10+
StakeNot required
Validator41% pool

Score miners on recall@K, latency, and proof rate. Set weights on-chain.

RAM8 GB minimum
Storage20 GB SSD
StakeTAO required
UptimeAlways-on
BuilderFree · testnet

Integrate Engram as your vector store using the Python SDK or CLI.

Installpip install engram-subnet
ModelsAny embedding model
AccessFree during testnet
LangPython 3.10+
quickstart.sh
git clone https://github.com/Dipraise1/-Engram-
cd engram
pip install -e ".[miner]"
python neurons/miner.py --wallet.name mywallet --netuid 42
Engram

The future of AI memory
is decentralized.

open source · bittensor subnet · testnet active · v0.1

Open Dashboard View Source