Summary

SUMA Memory is a persistent semantic knowledge graph for AI assistants, distributed as an npm package (suma-mcp-proxy).

What it does

• Gives Claude (and other MCP-compatible AI) permanent, searchable memory that survives session boundaries
• Uses the K-WIL Gravity Well Algorithm — a retrieval system that scores nodes by Recency × Harmonic Weight × Semantic Similarity × Persona Profile
• At 180K+ tokens of stored knowledge, a single suma_search retrieves the exact 800 tokens the AI needs — no hallucination, no context window overflow
• Supports Claude Code, Cursor, Windsurf via .mcp.json

Why K-WIL vs standard RAG

Standard RAG does cosine similarity ranking. K-WIL adds:

• Recency decay — recent facts rank higher than stale ones
• Harmonic weight — well-connected entities rank higher (log-flattened to prevent black holes)
• Persona profiles — per-industry weight adjustments (developer, enterprise, healthcare, government)

Result: retrieval precision, not just recall.

Install

One-line setup — no global install required:

{
  "mcpServers": {
    "suma-memory": {
      "command": "npx",
      "args": ["suma-mcp-proxy", "--key=sk_live_YOUR_KEY"]
    }
  }
}

Links

• npm: https://www.npmjs.com/package/suma-mcp-proxy (v1.3.4)
• Source: https://github.com/sumanaddanki/suma-mcp-proxy
• Dashboard: https://sumapro.quadframe.work

Tools exposed

Checklist

• Package published on npm: suma-mcp-proxy@1.3.4
• Source repo is public: github.com/sumanaddanki/suma-mcp-proxy
• Server tested with MCP Inspector (npx @modelcontextprotocol/inspector)
• Missing API key returns readable error (does not crash)
• JSON Schema uses additionalProperties: false
• Works with Claude Code, Cursor, Windsurf