Frontier models keep getting smarter. We're building the brain they don't have.
Today's LLMs are powerful, stateless, and forgetful. Every conversation starts at zero. Every session re-learns what the last one already knew. Every agent built on top of one inherits the same gap.
Neruva is the substrate that fixes that — persistent memory plus deterministic reasoning, exposed via the same MCP your agent already speaks, deployable in one line, owned by you in a portable file.
We didn't want to build it. We needed it.
The first version of Neruva was an internal tool — a way to make our own AI agent loop survive across sessions without re-pasting ten thousand tokens of context every time.
We tried what everyone tries first: stuff transcripts back into the prompt. It was slow, expensive, and the agent still forgot everything that wasn't in the immediate window. We tried bolting on a vector DB. It was better at recall, worse at reasoning, and we ended up writing a memory schema by hand.
We tried key-value stores, hybrid retrieval, graph databases. Every layer fixed one thing and broke another. None of them gave us causal reasoning, episodic memory, or deterministic replay — the parts that actually let an agent learn from experience instead of just searching its own logs.
So we built Neruva. One substrate, every kind of memory the agent needs, one billing ledger, sub-100ms recall. It works for our agent. It works for yours.
The principles we keep coming back to.
Your data is yours.
One .neruva file carries every layer of your agent's substrate — records, knowledge graphs, causal models, episode stores, induced rules, K-gram learning state, theory-of-mind shards. Bit-identical replay on import, tested end-to-end. The format is documented and open. If we ever go away, you walk out the door with everything.
Bring your own LLM.
We don't compete with Claude, GPT, or open-weights models. We make them better at remembering. Pair us with whatever frontier model fits your stack.
Deterministic from a seed.
Every operation is replayable bit-for-bit. The same query against the same data returns the same answer, every time. The audit story regulators want, the debug story your team needs.
No research-paper download.
Capabilities ship in the public API or they don't count. If you can't call it from MCP in one line, we haven't shipped it yet.
Trust by transparency, not certifications.
We don't have SOC 2 yet. We're early. What we do have: the source code your agent talks to, public latency claims you can verify, deterministic replay you can audit yourself. Certifications come when paying customers need them.
Sub-100ms or it isn't real.
Memory that takes a full second to recall is memory the agent stops using. The substrate runs at integer-multiply / sign / cosine speeds. The latency floor is the network, not us.
Small, focused, building in the open.
Neruva is the product. Clouthier Simulation Labs is the Ontario-based company behind it. We're small. We ship a lot. We answer support email ourselves.
Founded by Kyle Clouthier in 2026. Built on the assumption that the next decade of AI is going to be about giving these systems durable memory and verifiable reasoning, not just larger context windows. Funded by the founder; runway is real, decisions don't need a committee.
If you want to talk — info@neruva.io.
Want to see what an agent that remembers feels like?
Thirty seconds, no card, every cognitive primitive on the free tier.