@natiwo/spectrum-mcp

MCP server for Spectrum — give your AI persistent memory.

Works with Claude, Gemini, Codex, and any MCP-compatible client.

What This Does

Your AI forgets everything between sessions. This MCP server connects it to Spectrum, a local SQLite-based semantic memory system. Your AI can save context, recall decisions, and search by meaning across sessions.

You: "Remember that we use Tailwind 4 with the Catalyst preset"
AI:  [saves to Spectrum]

--- next day, new session ---

You: "What CSS framework do we use?"
AI:  [searches Spectrum] "You use Tailwind 4 with the Catalyst preset."

No cloud. No API keys. Everything on your machine.

Setup

Claude Code

Add to ~/.claude.json:

{
  "mcpServers": {
    "spectrum": {
      "command": "npx",
      "args": ["-y", "@natiwo/spectrum-mcp"]
    }
  }
}

Claude Desktop

Add to claude_desktop_config.json:

{
  "mcpServers": {
    "spectrum": {
      "command": "npx",
      "args": ["-y", "@natiwo/spectrum-mcp"]
    }
  }
}

Global install (faster startup)

npm install -g @natiwo/spectrum-mcp

Then use spectrum-mcp instead of npx:

{
  "mcpServers": {
    "spectrum": {
      "command": "spectrum-mcp"
    }
  }
}

Any MCP client

Spectrum MCP uses stdio transport. Point your client to the spectrum-mcp binary or npx @natiwo/spectrum-mcp.

Tools

The server exposes 7 tools:

| Tool | Description | |------|-------------| | memory_save | Save or update a memory (scope, key, value, optional tags) | | memory_get | Retrieve a specific memory by scope and key | | memory_list | List memories in a scope (with optional prefix filter) | | memory_delete | Delete a specific memory | | memory_search | Keyword search across keys and values | | memory_search_semantic | Search by meaning using embeddings | | memory_list_scopes | List all scopes with memory count |

Example tool calls

memory_save(scope: "project:my-app", key: "stack", value: "Next.js 15, PostgreSQL, Redis", tags: ["architecture"])
memory_get(scope: "project:my-app", key: "stack")
memory_search(query: "database", scope: "project:my-app")
memory_search_semantic(query: "which database did we choose?")
memory_list(scope: "project:my-app", prefix: "config")
memory_list_scopes()
memory_delete(scope: "project:my-app", key: "old-decision")

Scopes

Organize memories by context:

| Scope pattern | Use | |---------------|-----| | user | Personal preferences, coding style | | project:<name> | Project-specific architecture, decisions | | session | Session summaries, continuity | | team:<name> | Shared team knowledge | | debug | What broke, what fixed it |

Semantic Search

For search by meaning (not just keywords), install the optional dependency:

npm install -g @huggingface/transformers

The embedding model (all-MiniLM-L6-v2, 384D, ONNX) downloads on first use and runs locally. No API. No cloud. No latency.

Without it, memory_search_semantic falls back to keyword search.

Storage

All data in ~/.spectrum/spectrum.db (SQLite, WAL mode). One file. Copy it, back it up, sync it between machines.

Custom path:

{
  "mcpServers": {
    "spectrum": {
      "command": "spectrum-mcp",
      "env": {
        "SPECTRUM_DB_PATH": "/path/to/custom.db"
      }
    }
  }
}

Ecosystem

| Package | What | |---------|------| | @natiwo/spectrum | Core library | | @natiwo/spectrum-cli | CLI tool (spm) | | @natiwo/spectrum-mcp | MCP server (you are here) |

Documentation

Full documentation, API reference, and architecture: github.com/Natiwo/spectrum

License

MIT