Pinecone announced it.
Neruva shipped it.
Pinecone's "knowledge infrastructure for agents" (Nexus + KnowQL) launched May 2026 as Early Access only -- gated to financial services, healthcare, legal, and enterprise SaaS sectors via an application form. Your pcsk_... key doesn't unlock it.
Neruva ships the substrate today. Open API, real benchmarks, sign up in 60 seconds, $5 free credit, no application form.
Same frame. Different substrate.
| Capability | Pinecone Nexus | Neruva |
|---|---|---|
| Availability | Early Access only -- application + sector qualification | GA today, sign up in 60 sec |
| Typed records (kind / tags / ts / meta) | Marketplace apps + per-field typed outputs | Open Records substrate -- records_ingest / records_query |
| Per-field citations + confidence | Yes (in compiled artifacts) | Yes (KG .confidence, analogy .cosine, causal .probability) |
| Knowledge graph with calibrated confidence | Implied via "artifacts" -- no public KG primitive | Sharded HD KG, /v1/hd/kg/*, 100% accuracy at 200 facts |
| Pearl do-operator (P(Y|do(X))) | Not offered | Two distinct endpoints -- observation vs intervention arithmetically separated |
| HD analogy (A:B::C:?) | Not offered | /v1/hd/analogy -- parallelogram completion in HD space |
| Portable single-file export | Hosted-only (BYOC enterprise = self-host stack, not portable file) | .neruva container -- one zip per namespace, atomic, point-in-time |
| Deterministic-from-seed replay | Versioned artifacts + audit trace | Bit-identical reruns of the same query (verified 20/20 on /benchmarks) |
| Open agent-memory protocol | KnowQL -- vendor-declared standard | Typed-record schema published in docs; any harness can adopt |
| Auto-record from Claude Code | Not offered | neruva-record-install -- one command, full session capture, secrets-redacted |
| Cost reduction claim | "Up to 90% token reduction" -- single internal financial-analysis test | ~3,125x cheaper per recall -- measured on /benchmarks against Opus 4.7 list rates |
| Pricing transparency | Builder $20/mo flat. Nexus pricing not disclosed. | Wallet-based metered, per-op rates published on /pricing |
| Distribution / ecosystem | LangChain, Box, LlamaIndex, Unstructured, Teradata endorsements | MCP install in Claude Code today; Cursor / Codex / Gemini hooks on roadmap |
The substrate is more than a context compiler.
P(Y | X=1) is what you observed. P(Y | do(X=1)) is what would happen if you forced it. Same logged worlds, two arithmetically distinct endpoints. The killer feature for principled action selection -- the substrate distinguishes co-occurrence from causation natively. Pinecone's artifacts can't do this -- it's a different math, not a different prompt.The honest part.
We're not pretending. Pinecone has things we don't.
Change one import.
Neruva's memory layer is Pinecone-shape compatible. Your existing upsert / query / delete code works unchanged. Then graduate to typed records, KG, causal, and analogy when you want them.
# before from pinecone import Pinecone # after from neruva import Pinecone
Or hit our REST directly with the Pinecone host swap: host="https://api.neruva.io" on the Pinecone client. Same JSON shapes. Egress is free. The full mapping is in /integrations.
~3,125x vs ~10x. One is reproducible.
"Up to 90% token reduction" -- per Pinecone's announcement.
Source: single internal financial-analysis benchmark per The New Stack. Methodology not public.
records_query at $0.000002 / call vs context-stuffing 5 KB at Opus 4.7 input rates ($0.00625 / turn).
Source: live measurements on /benchmarks. Reproducible with one Python file + your own API key.
Don't wait for the application form.
Sign up at app.neruva.io, get $5 free credit, and ship the substrate into your agent in 60 seconds. No sector qualification, no waitlist, no NDA.