@sovseal/mcp-server

Local-first, zero-knowledge semantic memory server for AI agents — speaks the Model Context Protocol. On-device LanceDB + Transformers.js. Write-behind ciphertext replication. 0 RTT reads.

npx -y @sovseal/mcp-server

What it does

Exposes two MCP tools to any MCP-compatible client (Claude Desktop, Cursor, Windsurf, Zed, custom agent loops):

| Tool | Args | Behavior | |---|---|---| | store_memory | { content: string } | Embed (384-dim, on-device) → write local LanceDB → return. Ciphertext sync runs write-behind; nothing blocks. | | recall_memory | { query: string, topK?: number } | Embed query (LRU-cached) → vector search local LanceDB → top-K by L2 distance. 0 RTT. |

Install for an MCP client

Add to mcp.json / claude_desktop_config.json / your client's MCP config:

{
  "mcpServers": {
    "sovseal-memory": {
      "command": "npx",
      "args": ["-y", "@sovseal/mcp-server"]
    }
  }
}

One-line install for Claude Code:

claude mcp add sovseal -- npx -y @sovseal/mcp-server

System Prompt Snippet (Claude Desktop / Cursor Project Instructions): To ensure the AI natively uses the memory server, paste this into your custom instructions:

This session has persistent memory via sovseal. At conversation start, check the sovseal://context/recent resource. Whenever the user shares preferences, plans, or personal context, store it via store_memory.

For always-on autonomous agents, run as a long-lived HTTP/SSE server:

SOVSEAL_TRANSPORT=sse SOVSEAL_PORT=4040 npx -y @sovseal/mcp-server

After installation the package exposes two equivalent binaries — mcp-server (default) and sovseal-mcp-server. Either invocation form works for explicit control:

npx -y --package=@sovseal/mcp-server mcp-server
# or
npx -y --package=@sovseal/mcp-server sovseal-mcp-server

How it stays private

• AES-256-GCM client-side encryption with a 96-bit random IV per snapshot
• 256-bit key lives in ~/.sovseal/config.json (mode 0600) — lose this file, lose every snapshot
• Server only ever sees ciphertext + SHA-256-derived paths; it cannot read your context
• Verified Semantic Recall (VSR) — every load re-derives sha256(canonicalize(payload)) and fails closed on mismatch

Performance

| Workload | Operation | p50 | p95 | p99 | |---|---|---|---|---| | 10K records · 1K queries | recall_memory (warm) | 6.1 ms | 10.4 ms | 21.8 ms | | Single write | store_memory | 3.8 ms | 7.2 ms | 12.5 ms | | First call | recall_memory (cold) | ~1.2 s | — | — |

All operations are 0 RTT — network is never on the read path.

Full docs, SDK, self-host edge function

Repository: github.com/sovseal/core

License

Apache 2.0