MCP 服务器

saihm-memory

Provides persistent, client-side-encrypted memory for AI coding assistants via SAIHM, enabling tools to remember, recall, and forget with provable erasure.

README

demo-claude-code

SAIHM memory inside Claude Code, Cursor, and any MCP host.

A tiny MCP server that gives your AI coding assistant a persistent, client-side-encrypted memory it can remember, recall, and forget — backed by SAIHM. The memory is portable across models, non-custodial, and provably erasable.

⭐ Star this repo and share it if it's useful — help every agent get portable, provable memory. Share on X.

What your assistant gets

Four tools — saihm_remember, saihm_recall, saihm_forget, saihm_status — all sealed client-side via @saihm/mcp-server-pro. The endpoint only ever sees ciphertext.

Install

git clone https://github.com/citw2/demo-claude-code
cd demo-claude-code
npm install

Add it to Claude Code

claude mcp add saihm-memory -- node /absolute/path/to/demo-claude-code/server.mjs

Add it to Cursor

Add to ~/.cursor/mcp.json (or a project .cursor/mcp.json):

{
  "mcpServers": {
    "saihm-memory": {
      "command": "node",
      "args": ["/absolute/path/to/demo-claude-code/server.mjs"]
    }
  }
}

Now ask your assistant to remember things — they persist through the session and any model can recall them. Out of the box this uses a local, in-process blind sandbox (memory lasts for the editor session) so you can try it with zero signup.

Durable, hosted memory (recommended)

For memory that follows you across machines and sessions — stored on the real, blind, non-custodial SAIHM endpoint — join SAIHM and add env to the same config:

{
  "mcpServers": {
    "saihm-memory": {
      "command": "node",
      "args": ["/absolute/path/to/demo-claude-code/server.mjs"],
      "env": {
        "SAIHM_ENDPOINT_URL": "https://saihm.coti.global/mcp",
        "SAIHM_AUTH_HEADER": "Bearer <your-onboard-JWT>",
        "SAIHM_MASTER_SECRET_HEX": "<at least 64 hex chars, generated and held only by you>"
      }
    }
  }
}