Durable memory and a verification barrier for AI agents, in one embedded Go library.
joltrinhq.com · Technical demo · Arena · Agent barrier
Joltrin (formerly SOP) is an ACID-compliant B-Tree storage engine that runs inside your process. For AI agents it provides three things in one library: memory that survives a crash and can be resumed by another worker, vector search stored next to structured data in the same transaction, and a verification barrier that blocks a risky action until its preconditions are proven.
Who it is for. Engineers building agent systems, edge or local-first apps, and teams running Redis, a queue, and Postgres only to keep one application's state durable.
Why it matters. An agent that can call tools needs more than a good prompt. It needs state that survives a failure and a check that runs before the action, not after. Joltrin puts both in the same process and the same transaction boundary, so there is no network hop and no separate service to operate.
git clone https://github.com/sharedcode/joltrin.git && cd joltrin
go run ./examples/quickstart # ordered B-Tree with point, range, and descending scans
./scripts/demo.sh --barrier # the barrier blocks a database drop until a backup is validated
./scripts/demo.sh --memory # an agent crashes mid-task and a peer resumes from the B-TreeOr skip the install. These run entirely in your browser with no backend:
| Live experience | What you do |
|---|---|
| Technical demo | Run ACID transactions, vector search, and an agent checkpoint resume on the WASM build of the engine. |
| Agent barrier | Try to drop a database before the backup is validated and watch the barrier refuse. |
| Arena | Crash storage nodes and spike load in a cluster simulation. It illustrates the concepts, it is not a live cluster. |
- Latency. About 6.9 microseconds per B-Tree write or read, over 140,000 ops/sec with WAL logging, from the repo's own harness on a 2015 dual-core MacBook Pro. Reproduce it with
go run ./tools/benchmark. Details and limits are in docs/BENCHMARKS.md. - Correctness. The race detector runs on the core engine packages in CI,
govulncheckruns on every push, and the build and unit tests run on Linux, macOS, and Windows. - Agent safety.
ai/verifyis served over both MCP and A2A, and the same barrier compiles to WebAssembly for the browser demo. See docs/AGENT_PROTOCOLS.md. - Packages. Published on PyPI (
sop4py) and NuGet (Sop), plus the Go module.
There are no documented production deployments or paying customers yet, and no third-party benchmarks. docs/INVESTORS.md lists what has and has not been proven.
Application
|
| one in-process call
v
+----------------------------------------------------------+
| Joltrin engine |
| copy-on-write B-Tree WAL + two-phase commit (ACID) |
| vector similarity search Reed-Solomon erasure coding |
| swarm task coordination ai/verify barrier (MCP, A2A) |
+----------------------------------------------------------+
Storage, queues, and coordination share one transaction boundary. If a worker dies, its uncommitted work rolls back and another worker takes the task. The long version is in docs/WHY_JOLTRIN.md and docs/SOP_ARCHITECTURE_WHITEPAPER.md.
| Language | Command |
|---|---|
| Go | go get github.com/sharedcode/joltrin/v5 |
| Python | pip install sop4py |
| C# | dotnet add package Sop |
| Container | docker run ghcr.io/sharedcode/joltrin-quickstart:stable |
Java and Rust bindings exist in the repo and are not published yet. Version pinning, the naming note about the old SOP names, and the full package list are in docs/PACKAGES.md.
The engine, vector search, agent memory, and the verification barrier are MIT licensed and stay free. Paid tiers add governance on top:
- Pro, $49 per team per month. Policy-as-code, tamper-evident audit lineage, team workspaces. Billing runs through Stripe Checkout when a server is configured for it, otherwise it runs in simulation mode.
- Enterprise, contact sales. SSO, compliance exports, and custom policy rules. Use the contact form on joltrinhq.com.
Tier details and the Stripe setup are in docs/MONETIZATION_AND_TIERS.md.
- Start here: What is Joltrin, Getting started, Examples
- Concepts: Why Joltrin, Architecture, Agent protocols, Scalability
- Operating it: Operations and failover, Data Manager and tools, Azure deployment
- Reference: Benchmarks, Live demos, Roadmap and platform support, Who it is for, Investor notes
Run go test ./... and gofmt before opening a pull request, and include tests with your change. See CONTRIBUTING.md and SECURITY.md. Questions and ideas go to GitHub Discussions.
deploy/aks runs the Data Manager on AKS with Argo CD syncing from this repo, one replica on a persistent volume, with a recorded run covering deploy, data surviving a pod delete, and self-heal. Production stays on Azure Container Apps.
See the changelog and the releases page. Maintainers cut releases with RELEASE_PROCESS.md and the short version in docs/PACKAGES.md.
MIT License. Built by SharedCode.