MCP 服务器

gnosem

Cross-vendor AI memory over MCP. Provides a persistent memory store with tools to write, search, list, forget, and supersede memories from any MCP-capable AI client.

README

Cross-vendor AI memory over MCP. One memory store. Every model. Claude, ChatGPT, Gemini, Kimi, Cursor, Windsurf — anything that speaks MCP.

<a href="https://gnosem.dev">gnosem.dev</a> · <a href="https://gnosem.dev/upgrade">pricing</a> · <a href="https://gnosem.dev/llms.txt">llms.txt</a>

What it is

Gnosem is a hosted Model Context Protocol server that gives each user one persistent memory store — readable and writeable from every MCP-capable AI client. Vendor-agnostic by design: a fact saved from Claude is visible in ChatGPT, Cursor, Windsurf, Zed, Kimi, and any other MCP client (or plain HTTP tool call) using the same API key.

The problem it solves: every AI vendor has its own proprietary memory. OpenAI's ChatGPT memory doesn't work in Claude. Anthropic's projects don't cross into GPT. Multi-model users retell context in every session. Gnosem is the neutral layer between vendors.

Tools exposed

memory_write — save a fact, preference, decision, or note. content, tags?, written_by? (provenance), session_id?, no_optimize?
memory_search — semantic search across your memories. query, k?, raw?
memory_list — list recent memories in reverse-chronological order. limit?, cursor?, raw?
memory_forget — soft-delete a memory by id
memory_supersede — replace a stale memory with a corrected one

AI-optimized storage

Long memories (>400 chars) are automatically compressed on write to a structured-facts form (Workers AI, llama-3.1-8b-instruct-fast) with a strict prompt: TOPIC=..., PROJECT=..., DECISION=..., STACK=..., PROBLEM=.... The reading LLM ingests the compact form by default — fewer tokens, same meaning. The raw prose is preserved and returned as content_raw; pass raw:true to invert. Pass no_optimize:true on write to skip compression entirely.

Compression is guarded — if the model output isn't actually shorter, gnosem falls back to storing raw only. Fail-open: any AI error still saves the memory.

Quickstart

npx gnosem-install

Auto-detects and configures every MCP-capable client on your machine (Claude Desktop, Claude Code, Cursor, Windsurf, Zed). Prompts for your API key and merges the gnosem entry alongside your existing servers. Restart the affected clients — the five memory tools appear immediately.

Get an API key at gnosem.dev, or via the API:

curl -sX POST https://gnosem.dev/signup \
  -H "Content-Type: application/json" \
  -d '{"email":"you@example.com"}'

The key is displayed once. Save it.

Manual config (any MCP client)

{
  "mcpServers": {
    "gnosem": {
      "url": "https://gnosem.dev/mcp",
      "headers": { "Authorization": "Bearer gn_your_api_key" }
    }
  }
}

Quickstart — any HTTP client

# Write
curl -sX POST https://gnosem.dev/mcp \
  -H "Authorization: Bearer gn_your_key" -H "Content-Type: application/json" \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/call","params":{"name":"memory_write","arguments":{"content":"I prefer Postgres for greenfield work.","written_by":"claude-code","tags":["preference"]}}}'

# Search
curl -sX POST https://gnosem.dev/mcp \
  -H "Authorization: Bearer gn_your_key" -H "Content-Type: application/json" \
  -d '{"jsonrpc":"2.0","id":2,"method":"tools/call","params":{"name":"memory_search","arguments":{"query":"database preference","k":5}}}'

Pricing

Plan	Price	Memories	API keys	Storage
Free	$0	200	1	shared
Pro	$9/mo or $90/yr	Unlimited	Unlimited	1 GB

Upgrade at gnosem.dev/upgrade.

Data model

Each memory row: id (uuid), content (raw, ≤8000 chars), content_optimized (structured facts, nullable), tags (json array), written_by (provenance — which model wrote it), session_id (opaque grouping), created_at, plus a semantic embedding (BGE-base-en-v1.5, 768-dim) stored in Vectorize.

Every memory tracks provenance so you can see which model contributed which fact. Memories are per-user isolated at both the D1 metadata layer (user_id filter on every query) and the Vectorize layer (metadata index on user_id).

Correction chains via memory_supersede: the old row is marked superseded (excluded from reads) and the new row references it. Soft-delete via memory_forget (excluded from reads and removed from the vector index).

Infrastructure

Cloudflare Workers — edge-hosted MCP server, no cold starts
D1 — metadata + auth
Vectorize — 768-dim embedding storage + cosine ANN search, per-user metadata filter
Workers AI — embedding generation (@cf/baai/bge-base-en-v1.5) and content optimization (@cf/meta/llama-3.1-8b-instruct-fast)
Stripe — subscription billing

Latency: sub-100ms for read, ~200-800ms for write (dominated by embedding + optimization inference).

AI discovery

The service is designed to be discoverable and citeable by LLMs:

/llms.txt — LLM-optimized site summary per llmstxt.org
/robots.txt — explicitly welcomes GPTBot, ClaudeBot, PerplexityBot, Google-Extended, and every major AI crawler
/sitemap.xml — standard sitemap
Schema.org JSON-LD (SoftwareApplication + FAQPage) embedded in the landing page

Self-hosting

The Worker source in this repo can be deployed to your own Cloudflare account. You'll need:

A Cloudflare account with Workers, D1, Vectorize, and Workers AI enabled
npx wrangler d1 create <name> — creates the D1 database. Update wrangler.jsonc with the returned id
npx wrangler vectorize create <name> --dimensions=768 --metric=cosine — creates the index
npx wrangler vectorize create-metadata-index <name> --property-name=user_id --type=string — required for per-user isolation
npx wrangler d1 execute <name> --remote --file=schema.sql then apply migrations/*.sql in order
npx wrangler secret put STRIPE_SECRET_KEY and STRIPE_WEBHOOK_SECRET (only needed if you want billing)
npx wrangler deploy

The hosted service (gnosem.dev) is the recommended way to use Gnosem — one API key, no infra to run, memories stay reachable when you switch machines.