记忆层级: working_memory (短期) → episodic_memory (情景) → persona (人格)
检索: hybrid 评分 (vector 50% + FTS5 30% + importance 20%)
图谱: subject-predicate-object 三元组 + 多跳遍历
蒸馏: 长期会话自动 consolidate，防止上下文爆炸

# 安装
pip install mnemosyne-mcp

# 初始化记忆库
mnemosyne-mcp init --path ~/.mnemosyne/data

# 启动 MCP Server (stdio)
mnemosyne-mcp serve

# 或 SSE 模式（本地回环）
mnemosyne-mcp serve --transport sse --port 8080

{
  "mcpServers": {
    "mnemosyne": {
      "command": "mnemosyne-mcp",
      "args": ["serve"]
    }
  }
}

from mnemosyne_mcp.client import MnemosyneClient

client = MnemosyneClient(path="~/.mnemosyne/data/mnemosyne.db")

# 批量写入（原子操作）
results = client.batch([
    {"action": "remember", "content": "用户偏好 Python", "scope": "global", "importance": 0.9},
    {"action": "triple_add", "subject": "user", "predicate": "prefers", "object": "Python"},
    {"action": "graph_link", "source_id": "...", "target_id": "...", "relationship": "references"},
])

from mnemosyne_mcp.client import MnemosyneClient

client = MnemosyneClient(path="./my_memory.db")

# 写入
mem_id = client.remember(
    content="用户喜欢简洁的代码风格",
    tags=["preference", "coding-style"],
    importance=0.8,
)

# 检索
results = client.recall("代码风格偏好", limit=5)

# 知识图谱
client.triple_add("user", "prefers", "Python", confidence=1.0)
graph = client.graph_query(seed_id=mem_id, max_hops=2)

# 记忆蒸馏
client.sleep(force=True)

mnemosyne-mcp/
├── src/mnemosyne_mcp/
│   ├── __init__.py      # 公共 API 导出
│   ├── server.py        # MCP stdio/SSE 服务器
│   ├── client.py        # 纯 Python 客户端（无需 MCP）
│   └── tools.py         # 工具注册 + 路由
├── tests/
│   ├── test_client.py
│   └── test_tools.py
└── examples/
    ├── basic_usage.py
    └── knowledge_graph.py