AI that runs on CPUs, memory and storage.
Semurg is software for ordinary CPU servers. We want to work with the clouds, chip makers and memory and storage makers whose hardware it runs on, and with their marketplace and licensing teams.
What Semurg is
Software that gives AI agents such as Claude Code and Cursor a memory that lasts, on CPU servers you run. Serving very large open models from storage is in development; it is not in a published release yet.
Why it matters
We think frontier AI is limited by memory, not arithmetic. Agents forget between sessions. Agent memory and data run on CPU cores, server memory and NVMe drives, and grow by adding servers.
What it does today
Agents remember decisions and facts across sessions, each with its source: 17 of 17 memory checks pass (release of 16 November 2026). The engine wrote 75.8 million 64-byte rows a second across two servers with 8 CPU cores each (release of 16 November 2026).
Not in a published release yet: serving very large models. We publish its speed only when a release measures it.
Some AI work can run on CPU servers.
Two kinds of AI work, and where each runs.
- Agent memory. Agents need to keep what they learn: decisions, facts, documents, and where each one came from. That is a database job. It runs on CPUs, server memory and drives, next to the data.
- Very large open models. Only a few experts of a mixture-of-experts model run for each word. Serving such models from storage is in development.
- A different question. A buyer can ask how many cores, memory channels and drives a workload needs.
Serving very large models from storage is in development and is not in a published release yet.
Why Semurg is relevant to you
Five kinds of team, one reason: every agent and every node uses CPU cores, server memory and drives.
Clouds and hyperscalers
Including regional, sovereign and bare-metal providers
- Agent workloads that can run on your CPU instances. Every Semurg node is a server the customer runs, on hardware they own or rent, and a cluster grows one node at a time.
- Data that stays in your customer's own account, on machines they control: often why regulated and sovereign buyers pick a cloud.
Working together: we would like to discuss co-engineering on your instance types, test capacity for our release measurements, a listing in your marketplace, and joint go-to-market for regulated and sovereign buyers.
CPU and chip makers
Processors and reference systems
- Semurg runs on CPUs only.
- Each release publishes how busy the CPU threads and drives were while it was measured. A new processor shows up as a published number.
- Agent memory runs on the CPU, next to the data.
Working together: we would like to discuss early access to new processors and reference systems, joint tuning measured on each release, reference designs, and joint publication of results with the method.
Memory and storage makers
DRAM and NVMe
- When data is larger than memory, speed depends on how fast memory and drives deliver bytes.
- A full scan read 27.5 GB/s from four drives, 90.5% of what they can deliver (release of 16 November 2026).
- A graph larger than its memory budget still walks at 627 million edge hops a second, reading from the drives (release of 16 November 2026).
Working together: we would like to discuss co-engineering on new memory and drive generations, test hardware measured on our release gate, and joint reference designs with published results.
Marketplace teams
Marketplaces, plugin directories and MCP registries
- Developers can install Semurg in one line and connect their agents over MCP. A listing would let them do that from your platform.
Working together: we would like to discuss a listing in your marketplace, plugin directory or MCP registry, a one-line install that runs on your platform, and joint launch material.
Licensing, OEM and embedded
Servers, storage systems, appliances and services
- Semurg can be licensed for your product, by agreement: a server line, a storage system, an appliance or a managed service.
- Your customers keep their data on the hardware you sold them.
Working together: licensing, OEM and embedding, by agreement. For reference, our public price for each node: the first node of every cluster is free forever; each additional node is about US$278 a year, or about US$1,390 once for the life of the machine.
What runs on your platform
Every Semurg node is a server: its cores, its memory and its drives. A cluster grows one node at a time.
What is measured, and when
Every figure here is on our Performance and Durability pages, with the release that measured it. The engine figures come from two servers with 8 CPU cores each and four drives, measured on the whole cluster at once.
| What | Result | Release |
|---|---|---|
| Write | 75.8 million rows a second (64-byte rows; the median of three runs; counted after the drives confirmed each row) | 16 November 2026 |
| Read | 14.4 million rows looked up by id a second | 16 November 2026 |
| Traverse | 982 million edge hops a second; 627 million a second with a memory budget smaller than the graph (the walks read from disk) | 16 November 2026 |
| Scan | 234 million rows a second; the drives delivered 27.5 GB/s, 90.5% of their ceiling | 16 November 2026 |
| Agent memory | 17 of 17 memory checks pass, on our own 2,000-document test set. A capability check, not a benchmark | 16 November 2026 |
| Crash test | Killed 21 times in 140 seconds under load: 32.2 million acknowledged writes, none lost; ready again in 0.66 s (median) | 29 October 2026 |
| Power cut | Writes acknowledged in the last seconds before a cut can be lost: at most 9.4 s in our test. Everything acknowledged earlier survived | 29 October 2026 |
The case, in three parts
Why now
- Agents are becoming everyday software, and each one needs a memory that outlives its session.
- We think frontier AI is limited by memory, not arithmetic.
- CPU servers, memory and NVMe drives grow by adding more of them.
Why Semurg
- It works today: agent memory over MCP, installed in one line, on machines the user runs.
- It is measured in public: the four operations are charted across 167 published releases, from 14 September to 16 November 2026, and each figure names its release.
- It runs on CPUs only.
- A graph larger than its memory budget still walks at 627 million edge hops a second, reading from the drives (release of 16 November 2026).
Why early
- Work with us on your hardware while the design is still open.
- If we agree, the first published measurements on your platform, with the method.
- A reference design and a listing, if we agree to them.
Ways to work together
| Way | What it looks like | Suits |
|---|---|---|
| Co-engineering | Our engineers and yours tune Semurg on your processors, memory, drives or instances. Results are measured and published per release. | Chip, memory and storage makers; clouds |
| Test capacity | Hardware or cloud capacity for our release measurements. Results on your platform are published with their release when you agree. | Everyone on this page |
| Marketplace listing | Semurg in your marketplace, plugin directory or MCP registry, with a one-line install on your platform. | Clouds; marketplace teams |
| Licensing, OEM and embedding | Semurg inside your server line, storage system, appliance or managed service. By agreement. | Licensing and OEM divisions |
| Joint go-to-market | Reference designs, joint results with the method, and guides written with you for your customers. | Everyone on this page |
| Strategic investment | Corporate venture teams of the companies whose platforms Semurg grows can talk to the founder about a strategic investment alongside a commercial relationship. | Corporate venture arms of clouds, chip, memory and storage makers |
Every way starts with a conversation about your platform.
Write to us
Write to semurg@semurg.org with your organisation, your role and your interest: co-engineering, test capacity, a marketplace listing, licensing or embedding, or joint go-to-market.
Partnership terms are agreed one to one, in writing.
Nothing on this page is an offer of securities or a solicitation of an offer to buy securities.
Claude Code and Cursor are trademarks of their owners. Semurg is not affiliated with or endorsed by them.