docs / start here

Documentation

Three local-first tools that compose into one stack: memory, tools, and cost control. Start with a 60-second install, then go as deep as you need.

The AI Trio is three open-source tools that fix three different failures of AI agents. They run locally, speak open protocols, and compose without a control plane.

Pick your path

The 60-second version

# Memory: store, then recall, in one SQLite file
npx -y @mem-os/sdk store "Prefers pnpm over npm" --db ./memos.db
npx -y @mem-os/sdk search "package manager" --db ./memos.db

# Tools: see every bundled MCP server
npx -y universal-mcp-toolkit list

# Cost: fold a long prompt and print the savings
npx -y llm-guardian optimize "your long context here"

How they compose

The trio is designed as a stack, not a bundle:

1. UMT is the transport layer. It writes the MCP config each harness understands, starts servers over stdio or HTTP, and can link a MemOS database so memory is available as an MCP tool.
2. MemOS is the memory layer. It stores the graph, retrieves context within a token budget, and exposes it over SDK, CLI, MCP, and HTTP.
3. Guardian is the cost layer. It folds long context, injects MemOS memory packs, and proxies Anthropic or OpenAI traffic so savings apply without code changes.

Read How the trio fits together for the contracts, ports, and data flow.

What "local-first" means here

• One file you own. MemOS is a SQLite database. Copy it, back it up, delete it.
• No telemetry. The core paths make no network calls.
• MIT licensed. No paid tier, no per-seat gate, no hosted control plane.
• Open protocols. MCP over stdio and Streamable HTTP; OpenAI- and Anthropic-compatible HTTP where it matters.

Where to go next