Scaffold a local-first AI-agent memory app in one command. No services, no API key, no account.
Quickstart · Documentation · Framework integrations · Changelog
Generates a working TypeScript starter powered by
actrone-memory: an agent with memory wired
up and running, with nothing to install but Node.
npm create actrone-app@latest my-agent
cd my-agent
npm install
npm startRequires Node.js 22 or newer. That is the whole prerequisite list.
Five files, no hidden config, nothing to delete before you start working:
my-agent/
├── src/agent.ts a runnable agent with memory wired up
├── package.json one dependency, actrone-memory; TypeScript 7 and tsx for development
├── tsconfig.json strict mode, ESM, ready for tsx
├── .gitignore node_modules, build output and .env files
└── README.md how to run it and where to go next
Run npm start and the agent stores a fact and a conversation turn, recalls what is relevant
to the next question, and prints it. The behaviour you came for is visible in the first thirty
seconds, not after you configure a vector database. Memory lives in the process by default, so
each run starts empty; Going to production shows how to keep it.
npm create actrone-app@latest my-agent -- --framework langgraphEvery project starts from the same agent. --framework adds the one command that wires memory
into that framework to the generated README and to the next steps the CLI prints:
npx actrone-memory add langgraph --write src/langgraph.tsIt writes one new self-contained file and prints the packages to install. The files come from
the examples actrone-memory type-checks against each framework in CI.
| Slug | Framework |
|---|---|
core |
None: the library on its own (default) |
vercel |
Vercel AI SDK |
langchain |
LangChain.js |
langgraph |
LangGraph.js |
mastra |
Mastra |
llamaindex |
LlamaIndex.TS |
openai-agents |
OpenAI Agents JS |
genkit |
Firebase Genkit |
voltagent |
VoltAgent |
claude-agent-sdk |
Claude Agent SDK |
cloudflare-agents |
Cloudflare Agents |
inngest-agentkit |
Inngest AgentKit |
--help prints the same list. An unknown slug fails immediately with the valid set rather
than scaffolding something broken.
This tool creates a new project. It refuses to write into a directory that is not empty.
If you already have a codebase, use the same command the starter uses, in your project:
npx actrone-memory add <framework>It prints a copy-paste recipe, or with --write <file> creates exactly one new
self-contained file. It never edits code you already wrote.
The starter is deliberately local-first, and it stays useful as-is. When you need more, there are two independent directions and neither requires a rewrite:
- Durable self-hosted. Pass stores to
MemoryManager.create(): Redis or Valkey for recent turns, and Qdrant or Postgres with pgvector for long-term memory. Same API, same result shapes. Installfastembedfor recall by meaning with a local model. - Governed and hosted. Actrone's hosted platform has not launched. When it does, its
ActroneMemoryManageris planned as a drop-in replacement forMemoryManager, adding PII tokenisation before inference, an audit trail and policy limits.
You are never required to take either step. actrone-memory is MIT and works standalone
indefinitely.
MIT. Free to use in any project, commercial or otherwise.