veclabs

AI agents are making decisions that affect real people. Their memory? A plain JSON blob on someone else's server. Anyone with database access can change it. The agent never knows. You can never prove it. We're fixing this

Check out the download stats for solvec by @veclabss on Pepy! pepy.tech/projects/solvec

most vector DBs treat memory like a black box. you write, you read, and you hope the agent remembered the right thing. we record a Merkle proof on Solana after every write. so you can prove exactly what the agent knew - and when.

Why most vector databases are slow: They're servers. Every query is a network round-trip. 50-200ms before you've even started searching. VecLabs runs in-process. No network. No serialization. No GC pauses. 4.7ms p99. That's the whole trick.

Phase 4 shipped. AES-256-GCM encrypted vector storage. Merkle root posted to Solana after every write. Memory survives restarts.

4.7ms p99 vector search at 100K vectors. OpenAI ada-002 embedding size (1536 dims). In-process. No network round-trip. No GC. Pinecone at the same scale: ~30ms p99. Pure Rust HNSW. Reproducible: github.com/veclabs/veclabs

Building VecLabs. Vector search for AI agents with cryptographic memory proof. Rust HNSW. WASM bridge. SHA-256 Merkle root on Solana after every write. 4.7ms p99 at 100K vectors. github.com/veclabs/veclabs