api.neruva.io · live·Free tier, no card

From typed memory to reasoning.

The layer that unlocks agentic operations.

Other memory vendors give your agent recall. We add the reasoning your LLM doesn't have natively — and supercharge every Claude / Cursor / LangChain workflow you already run.

  • Typed memory: records, 5 KG engines, episodic CBR — sub-100ms recall across sessions, projects, teammates.
  • Reasoning primitives no other memory vendor ships: Pearl do-operator, counterfactual rollouts, theory of mind, schema lifting, EFE planning, rule induction.
  • Code-graph across 8 languages: Python · JS · TS · Go · Rust · Java · Ruby · C/C++. Auto-fires on SessionStart, sub-ms structural queries, free in every tier.
  • One MCP install. Bring any LLM. Free tier, no card.
MCP server, Python & TypeScript SDKs, raw REST. Bring your own LLM. For Claude Code users →
What no other agent-memory vendor ships

Memory is the floor. Reasoning is the moat.

Six capabilities that move agents from “recall” to “actually reasoning over what they recalled.” The tick column is empty on the right for a reason — no one else has shipped these yet.

CapabilityNeruvaZepMem0Pinecone NexusAnthropic Memory
Persistent typed memory
Knowledge graph (multiple engines)5 enginesTemporal KG
Causal reasoning (Pearl do-operator)
Counterfactual rollouts
Theory of mind (nested belief)
Few-shot rule induction + persist
Episodic CBR + concept blending
Bit-identical replay from seed
One-line MCP install, any LLMClaude-only
Snapshot from public vendor docs, May 2026.See the full comparison →
Your agent
Claude Code / Cursor / LangChain / yours
MCP
Neruva substrate
Memory + reasoning
records · KG · causal · analogy · CBR · cognitive primitives
< 100ms
Recall
Decisions · facts · episodes · rules

One API key. One billing ledger. Sub-100ms p95 from a warm caller.

What changes for your agent

Same agent. Same model. Different morning.

Without Neruva
You
Help me fix that race condition in the worker queue.
Agent
Which worker queue? Could you share the file? What kind of race — is it a double-dequeue or a duplicate enqueue?
You
We talked about this Monday. The orders service. The fix was exponential backoff, not jitter.
Agent
I don't have context from previous sessions. Could you paste the relevant code so I can help?
With Neruva
You
Help me fix that race condition in the worker queue.
Agent
I see the decision from May 11 — orders service, exponential backoff (not jitter), Marco signed off. Two related mistakes are tagged "deploy" — want me to check those first?
You
Yes. And remind me what we decided about the retry cap.
Agent
Three retries, then dead-letter queue. Decision from May 7, after the incident with the billing webhook.
Per-recall cost

~3,000× cheaper
than re-pasting your conversation.

Context-stuffing Opus 4.7 at list price costs about $0.00625 per turn. A Neruva recall returns the same context for about $0.000002 — and arrives in sub-100ms with deterministic replay.

Opus 4.7 context-stuff$0.00625 / turnNeruva recall$0.000002 / call(barely visible at scale)3,000× gap

Source: Anthropic public list pricing for Opus 4.7 at typical 800-token recall window. See benchmarks for the full math.

Records
Typed event memory.
kind + tags + ts + meta + auto-embed. Hybrid OPB+cosine RRF retrieval. Compact, replace, forget — deterministic from a seed.
Knowledge graph
5 engines, one API.
Hadamard, OPB, multishard, 3-shard quorum, feature-bundle. Scales from a handful of facts to hundreds of thousands per matrix. N-hop derive in a single matmul.
Causal + analogy
Pearl do-operator, HD parallelogram.
Observation vs intervention. Cross-domain analogy retrieval. Stuff every other vendor calls research — shipping today.
Portable
.neruva container.
One file. Every layer — records, KGs, causal models, episode stores, rules, K-gram, theory-of-mind. Bit-identical replay on import, tested end-to-end. Your data is yours — no lock-in, ever.
🧠

Claude Code agent that remembers

Drop neruva-record into your Claude Code hook. Every session, every decision, every mistake — recallable across projects.

🗺️

Code-graph across 8 languages

Python · JS · TS · Go · Rust · Java · Ruby · C/C++. Auto-indexes on SessionStart; agent asks 'who calls bind()' in microseconds, not after greppping ten files. Sub-ms, $0/call, free in every tier.

📦

Typed agentic events

decisions, mistakes, tool_calls, llm_turns — auto-embedded with one federated agent_remember call. Smart routing across records + KG.

🔍

Cross-project semantic recall

Sub-100ms search over months of agent transcripts. Customizable context blocks: entity-grounded, deterministic, $0/call (vs Zep per-call).

🧬

Knowledge graph that scales

Five engines including a 3-shard quorum mode that surfaces adversarial shard tampering. Multi-hop derivation in one call.

🎯

Causal reasoning, not RAG

Pearl do-operator over your memory. "What if X had happened" returns a different answer than "X happened" — arithmetically distinct, not a paraphrase.

🎭

Theory of Mind, nested belief

Alice believes Bob believes Carol believes X. Position-tagged binding keeps depths separable so each layer of "believes that" is independently queryable.

🔮

Counterfactual rollouts + EFE planning

What if at step k the agent had taken action a' instead? Replayable, deterministic. KL-optimal plan selection on caller-owned dynamics.

🧩

Few-shot rule induction + continual K-gram

Learn a named transformation rule from a few demos. Sharded K-gram next-token predictor with integer-add accumulation — no catastrophic forgetting.

♻️

Deterministic replay from seed

Bit-identical reruns. Auditable, exportable, regulator-friendly. The compliance story nobody else ships.

🛡️

Secrets redaction client-side

14-pattern leak guard before bytes leave your machine. Recalled context wrapped in injection-guard tags. Built-in.

Built privacy-first

Your data stays yours.

·14-pattern secrets redaction client-side — keys never leave your machine.
·Recalled context wrapped in injection-guard tags — treat-as="data-only" by default.
·Export every substrate layer as one .neruva file — bit-identical replay on import, tested.
·GDPR tombstone-compaction at every flush. Deleted means deleted, not flagged.
How it works

Get a free API key

Sign in at neruva.io/dashboard. Free tier: 10k records, 100 recalls/day, no card required.

Install the MCP server

One line in your Claude Code / Cursor config. Or pip install neruva-mcpfor Python, @neruva/mcp on npm.

Your agent remembers

52 tools across records, 5 KG engines, causal, analogy, CBR. Recall is sub-100ms. Build whatever agent on top.

Why Neruva

What other memory vendors can't do.

Zep wins on temporal KGs. Pinecone Nexus is a context compiler. Anthropic Memory is key-value persistence. Neruva is the only substrate that stacks KG + causal + analogy + episodic in one API, sub-100ms, deterministic from a seed.

vs Zep / Mem0 / Letta

They store temporal facts. Neruva adds Pearl do-operator (causal), HD analogy parallelogram, and CBR episodes — reasoning, not just recall. Export as one portable .neruva file. No lock-in.

vs Pinecone / vector DBs

They're a layer. Neruva is a substrate — records + KG + causal + analogy + episodic, one API key, one billing ledger. Hybrid entity-grounded + cosine retrieval beats cosine alone on multi-hop questions.

vs Anthropic Memory tool

That's key-value persistence inside one vendor's SDK. Neruva is vendor-neutral memory infrastructure — works with Claude, OpenAI, open-weights, your own. Bring your LLM, we handle the brain.

vs context-stuffing

~3,000× cheaper per recall vs cramming the conversation back into the prompt. Sub-100ms p95 (mostly network). Server-side compute is microseconds — every op is integer multiply / sign / cosine.

Stop bolting on a vector DB.
Drop in the substrate.

Free tier: 10k records, 100 recalls/day, no card. Pro $20/mo unlocks 1M records + unlimited cognitive primitives. See pricing →