How Neruva compares
Honest read against the two vendors developers compare us to most: Pinecone (vector database, going broader via Nexus) and Zep (agent-memory layer with temporal KG). We're stronger on capability breadth and price; they're stronger on brand and compliance certifications. Here's the apples-to-apples.
Pricing
Verify current pricing at pinecone.io/pricing · getzep.com/pricing · neruva.io/pricing.
| Tier | Pinecone | Zep | Neruva |
|---|---|---|---|
| Free tier | 2 GB · 2M write · 1M read · 5 indexes | 1,000 credits/mo (~350 KB total) | 10k records · 100 recalls/day · 30 cognitive primitive calls/day |
| Entry paid | Builder $20/mo (10 GB, 5M write) | Flex $125/mo (50k credits ≈ 17 MB) | Pro $20/mo (1M records · unlimited cognitive primitives) |
| Mid | Standard $50+/mo (pay-per-use: $4-18 per M units) | Flex Plus $375/mo (200k credits ≈ 70 MB) | Startup $149/mo (10M records · multi-tenant federation) |
| Enterprise | $500+/mo (all clouds + regions) | Custom (SOC 2, HIPAA at this tier) | $499+/mo (deterministic replay export · audit log) |
Pay-per-use note: Pinecone Standard at 10M reads/month ≈ $160-180 just for reads. Zep's "1 credit = 350 bytes" means free tier covers ~350 KB of memory before upgrade. Neruva's caps are call-count, not bytes.
Capabilities — and what each enables for your agent
✓ = shipped · ~ = partial / early access · — = not available. The right-most column is what your agent can actually ask once the capability is wired in.
| Capability | Pinecone | Zep | Neruva | Your agent can now answer… |
|---|---|---|---|---|
| Vector retrieval | ✓ | ✓ | ✓ | "What did we discuss about Acme last week?" |
| Hybrid (vector + lexical) | ✓ | ✓ | ✓ | "Find any mention of the SOC2 audit findings" |
| Knowledge graph | ~ | ✓ | ✓ | "List every customer Sarah owns who hasn't renewed yet" |
| Causal queries (Pearl do-operator) | — | — | ✓ | "Did the price cut cause the renewal, or were they going to renew anyway?" |
| HD analogy (a:b::c:?) | — | — | ✓ | "Acme is to Globex as Initech is to ?" |
| Episodic case-based reasoning | — | — | ✓ | "Show me the 3 most similar past customer-churn cases" |
| Theory of mind (depth-unlimited) | — | — | ✓ | "Does Alice know that Bob knows about the layoff?" |
| Counterfactual rollouts | — | — | ✓ | "If we had offered 20% discount instead of 15%, would Globex have renewed?" |
| Active inference planning | — | — | ✓ | "Score these three 3-step plans by KL distance to the goal state" |
| Schema lifting (analogical pattern match) | — | — | ✓ | "This instance has the same (role, filler) structure as our churn template" |
| Continual learning (provable no-forgetting) | — | — | ✓ | "Train on this week's new ticket categories without losing last quarter's" |
| Hierarchical chunking (L^K recursive) | — | — | ✓ | "Decode the 4th paragraph of the 3rd section of the contract in one op" |
| Few-shot rule induction | — | — | ✓ | "Here are 3 demos — figure out the rule and apply it to the new case" |
| Tenant-specific PII rules | — | — | ✓ | Train rule from 5 examples → auto-redact your EMP-12345-AB codes |
| Typed-shape context output | ~ | — | ✓ | Return as `{tier: str, renewal_date: str, employees: int}` with per-field cites |
| Bit-identical replay from seed | — | — | ✓ | "Re-run last Tuesday's session to audit exactly what the agent saw" |
| Open reproducible benchmark | — | — | ✓ | Hi-ToM 84.58%, 99.58% byte-identical replay (open harness) |
| MCP-native (Claude Code, Cursor, Codex, Gemini CLI) | — | — | ✓ | One-line install: `npx -y @neruva/mcp@latest` |
| Auto-record SDK (Claude Code hook) | — | — | ✓ | Every chat turn + tool call lands in your substrate, secrets-redacted |
| Framework adapters (LangChain / LangGraph / CrewAI) | ~ | ~ | ✓ | Drop-in `BaseChatMessageHistory`, `BaseCheckpointSaver`, etc. |
Code-graph: how Neruva compares to code-context tools
Different category from vector DBs and memory layers. If your agent needs to know “who calls X” or “what class is Y defined in” sub-millisecond and across sessions, the comparison set is Sourcegraph / SCIP / LSP and the AI-coding incumbents. Same substrate as the cognitive primitives above — one API key, one billing ledger, one portable .neruva file.
Hosted service, per-seat enterprise pricing, web UX. Strong on cross-repo search. But: separate billing from your AI stack, no MCP integration, no agent-memory pairing, no causal / ToM / replay primitives. Neruva runs in-process via MCP and sits in the same engine as your records + KG + episode store.
Per-process, dies when the editor closes, scoped to a single open project. No persistence across sessions, no cross-codebase queries, no portable export. Neruva's index lives in the substrate — index once, query forever from any agent, hand the .neruva file to a teammate.
They re-grep + re-read every session: O(tokens × codebase) burned every turn for the same information. Neruva indexes once at SessionStart (opt-in via NERUVA_AUTO_CODE_INDEX=1) and answers sub-ms forever after. Substrate-side cost: $0/call.
You'd build a multi-language extractor, a triple schema, a storage layer, a query layer, MCP wiring, and pay several service bills. Neruva ships that under one API key today across 8 languages: Python · JavaScript · TypeScript · Go · Rust · Java · Ruby · C / C++.
Measured on real codebases: 99%+ macro recall@1 across calls / callers / defined_in_class / defined_in_module / imports, p95 query latency in the low milliseconds, $0/call on the substrate side. Available now across 8 languages (Python, JS, TS, Go, Rust, Java, Ruby, C/C++) via code_kg_callees / callers / class_of / module_of / imports (5 MCP tools). Included in every tier — code-graph storage does not consume the records cap. See /benchmarks for the open harness.
Which tool fits which job
Different stacks are built for different problems. The quick read.
Pinecone
- Vector search at scale, multi-cloud, multi-region.
- Teams with their own agent / memory layer who just need fast retrieval underneath.
- Integration with their managed inference, reranking, and KnowQL roadmap.
Zep
- Temporal knowledge graphs with automatic fact extraction as the primary need.
- Managed single-vendor memory with one-integration setup.
- Stamped SOC 2 / HIPAA today at Enterprise pricing.
Neruva
- Agents that need to think, not just retrieve. Theory of mind, counterfactuals, causal do-operator, schema lifting, EFE planning, rule induction — primitives no vector DB or memory layer ships.
- Compliance through replay. Bit-identical deterministic recall from a seed — auditable, exportable, three months later under audit.
- MCP-native, one-line install. Drop into Claude Code, Cursor, Codex, Gemini CLI, LangChain, CrewAI — same API key everywhere.
- Code-graph across 8 languages. Python · JS · TS · Go · Rust · Java · Ruby · C/C++ all index into the same substrate. Auto-fires on SessionStart, sub-ms structural queries (callers, callees, classes, imports), $0/call. Included in every tier.
- Builder-friendly economics. Free tier ships every engine. $20/mo unlocks unlimited cognitive primitives. ~3,000× cheaper per recall than context-stuffing Opus 4.7.
Where we're investing in 2026.
We ship in public. Here's the roadmap for the surfaces customers most commonly ask about.
Cloud Run region-pinned deployment for customers with EU data-residency needs. Spin-up time: about a week.
Read-only / namespace-scoped / op-class-scoped API keys for enterprise key hygiene.
Audit in progress with a Big-4 firm. Same controls already in production; the certification follows.
For customers building on top of regulated health workflows. Same substrate, signed BAA.
Multi-region failover and additional Cloud Run regions on the way. We start with the regions our paying customers ask for.
MCP hosts and framework adapters land continuously. If your tool isn't on /integrations yet, ask — most can ship in a week.