MCP 服务器

engram-mcp

Persistent semantic memory for AI agents. SQLite-backed, local-first, zero config. Semantic search via Ollama embeddings with keyword fallback. Tools: remember, recall, history, forget, stats.

README

@cartisien/engram-mcp

Persistent semantic memory for AI agents — MCP server powered by @cartisien/engram

Give any MCP-compatible AI client (Claude Desktop, Cursor, Windsurf) persistent memory that survives across sessions.

npx -y @cartisien/engram-mcp

What it does

Exposes 5 tools to any MCP client:

Tool	Description
`remember`	Store a memory with automatic embedding
`recall`	Semantic search across stored memories
`history`	Recent conversation history
`forget`	Delete one memory, a session, or entries before a date
`stats`	Memory statistics for a session

Memories are stored in SQLite. Semantic search uses local Ollama embeddings (nomic-embed-text) — no API key, no cloud. Falls back to keyword search if Ollama isn't available.

Quick Start

Claude Desktop

Add to ~/Library/Application Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "engram": {
      "command": "npx",
      "args": ["-y", "@cartisien/engram-mcp"],
      "env": {
        "ENGRAM_DB": "~/.engram/memory.db"
      }
    }
  }
}

Restart Claude Desktop. You'll see remember, recall, history, forget, and stats available as tools.

Cursor / Windsurf

Add to your MCP config:

{
  "mcpServers": {
    "engram": {
      "command": "npx",
      "args": ["-y", "@cartisien/engram-mcp"]
    }
  }
}

Configuration

Env Var	Default	Description
`ENGRAM_DB`	`~/.engram/memory.db`	SQLite database path
`ENGRAM_EMBEDDING_URL`	`http://localhost:11434`	Ollama base URL for embeddings

Local Embeddings (Recommended)

Install Ollama and pull the embedding model:

ollama pull nomic-embed-text

Semantic search activates automatically. Without Ollama, keyword search is used.

Example Usage

Once connected, your agent can:

remember(sessionId="myagent", content="User prefers TypeScript over JavaScript", role="user")

recall(sessionId="myagent", query="what are the user's coding preferences?", limit=5)
# Returns: [{ content: "User prefers TypeScript...", similarity: 0.82 }, ...]

history(sessionId="myagent", limit=10)

stats(sessionId="myagent")
# { total: 42, byRole: { user: 20, assistant: 22 }, withEmbeddings: 42 }