Operator Manual

Technical integration guides for the SEAM runtime.

  • Quickstart
  • MCP Integration
  • CLI Reference
  • Canonical Store

Getting Started

SEAM is designed to run locally as a binary. It binds to a directory to act as its storage root.

# Initialize a new SEAM root
seam init ./my-agent-memory

# Start the runtime server and dashboard
seam serve --port 8765

Ingesting Source Data

You can compile raw files into MIRL records using the CLI compiler.

seam ingest ./docs/**/*.md --tag "documentation"

Model Context Protocol

SEAM natively supports MCP over stdio, allowing direct integration into Claude, Cursor, and any MCP-compatible agent.

Add the following to your agent's config:

{
  "mcpServers": {
    "seam": {
      "command": "seam",
      "args": ["mcp", "--root", "./my-agent-memory"]
    }
  }
}

This exposes tools like seam_search, seam_retrieve, and seam_context_pack to the LLM.

CLI Commands

The seam CLI is the primary operator surface.

seam init [path]        # Initialize a new canonical store
seam ingest [files]     # Compile and index sources
seam serve              # Start REST API & Dashboard
seam mcp                # Start MCP stdio server
seam doctor             # Verify integrity and rebuild indexes
seam query "string"     # Test retrieval logic

The Canonical Store

SEAM enforces a strict engineering invariant: SQLite is the canonical source of truth.

When you ingest a file, it is compiled into a Machine Intermediate Representation Language (MIRL) record and saved durably to SQLite.

All vector stores (HNSW, FAISS, etc.) are treated as derived acceleration layers. If a vector index is corrupted, you simply run seam doctor --rebuild to reconstruct it perfectly from the SQLite truth.