🧠 coolmem — Cross-Agent Memory Fabric

A privacy-first MCP server that lets different AI agents (Cursor, Claude Desktop, etc.) share a persistent Knowledge Ledger — powered by local vector embeddings and SQLite.

No API keys. No data leaves your machine.

Architecture at a Glance

Agent A (Cursor)                 Agent B (Claude)
     │                                │
     │  store_memory(content, cat)    │  search_memories(query)
     ▼                                ▼
┌──────────────  MCP  stdio  ─────────────────┐
│               coolmem server                │
│  ┌────────────────────────────────────────┐ │
│  │  @xenova/transformers                  │ │
│  │  all-MiniLM-L6-v2 (384-dim, local CPU) │ │
│  └────────────────┬───────────────────────┘ │
│                   │  embed()                 │
│  ┌────────────────▼───────────────────────┐ │
│  │  SQLite  (coolmem.db)                  │ │
│  │  Memory { id, content, category,       │ │
│  │           embedding JSON, createdAt }  │ │
│  └────────────────────────────────────────┘ │
└─────────────────────────────────────────────┘

Quick Start

# 1. Install dependencies
npm install

# 2. Generate Prisma client + create the SQLite DB
npx prisma generate
npx prisma db push

# 3. Build
npm run build

# 4. Run (for manual testing)
node dist/index.js

The server communicates over stdio — you don't call it directly. Hook it into Claude Desktop or Cursor via the config below.

Connecting to Claude Desktop

Open (or create) your claude_desktop_config.json:

• macOS / Linux: ~/.config/claude/claude_desktop_config.json
• Windows: %APPDATA%\Claude\claude_desktop_config.json

Add the following block (adjust the path to match where you cloned this repo):

{
  "mcpServers": {
    "coolmem": {
      "command": "node",
      "args": ["C:/Users/Dell/Desktop/YCU/coolmem/dist/index.js"],
      "env": {
        "DATABASE_URL": "file:C:/Users/Dell/Desktop/YCU/coolmem/coolmem.db"
      }
    }
  }
}

Tip: After editing the config, restart Claude Desktop. You should see coolmem appear in the Connected Servers list.

Connecting to Cursor

In Cursor → Settings → MCP → Add New MCP Server:

| Field | Value | |---------|------------------------------------------------------------| | Name | coolmem | | Command | node | | Args | C:/Users/Dell/Desktop/YCU/coolmem/dist/index.js | | Env | DATABASE_URL=file:C:/Users/Dell/Desktop/YCU/coolmem/coolmem.db |

Available MCP Primitives

🔧 Tool: store_memory

Store a lesson learned, architectural decision, or any project knowledge.

| Parameter | Type | Description | |------------|--------|------------------------------------------------------| | content | string | The knowledge text to persist | | category | string | Label: architecture, bug-fix, decision, etc. |

Example agent prompt:

Use the store_memory tool. content: "We switched from REST to tRPC because it eliminates manual type duplication between client and server." category: "architecture"

🔧 Tool: search_memories

Semantically search the ledger with a natural-language query.

| Parameter | Type | Description | |-----------|--------|-------------------------------| | query | string | What you want to know about |

Returns the top 3 most relevant memories with cosine-similarity scores.

Example agent prompt:

Use search_memories. query: "Why did we change our API layer?"

📄 Resource: project_timeline

URI: memory://coolmem/timeline

A chronological list of the last 10 things the agents stored — useful for onboarding a fresh agent to the current state of a project.

Project Structure

coolmem/
├── src/
│   ├── index.ts          # MCP server — tools & resource definitions
│   ├── embeddings.ts     # @xenova/transformers wrapper (lazy singleton)
│   ├── similarity.ts     # Cosine similarity function
│   └── db.ts             # Prisma client singleton
├── prisma/
│   └── schema.prisma     # Memory model → SQLite
├── dist/                 # Compiled output (git-ignored)
├── models/               # Cached model weights (auto-downloaded, git-ignored)
├── coolmem.db            # SQLite database (git-ignored)
├── .env                  # DATABASE_URL
├── package.json
└── tsconfig.json

First-Run Note on Model Download

On the very first call to either tool, @xenova/transformers will download all-MiniLM-L6-v2 (~22 MB) into the ./models/ folder. Subsequent calls use the cached weights instantly. This is a one-time cost.

Extending This Server

| Want to add… | How | |-------------------------------|------------------------------------------------------------------| | Delete a memory | Add a delete_memory(id) tool via server.tool() | | List memories by category | Add a list_memories(category) tool | | More search results | Change .slice(0, 3) to .slice(0, N) in index.ts | | Faster vector search at scale | Swap the JS cosine loop for sqlite-vss or usearch | | Different model | Change MODEL_NAME in embeddings.ts |