MongoDB 作LangGraph的长短期记忆 / ai #71

先简单回顾下，长短期记忆就是构建大模型的上下文，以补充其额外的信息，如下所示：

messages=[
    {"role": "user", "content": "你好，我叫张三"},
    {"role": "assistant", "content": "你好，张三，很高兴认识你。"},
    {"role": "user", "content": "你知道我叫什么吗？"}
  ]

Agent做记忆功能其实就是在封装这一层，用临时的字典做短期记忆，或是用数据库做长期记忆。可以采用MongoDB作为Agent长久记忆的存储空间。与LangGraph结合时，有两种写法.

uv add langgraph-checkpoint-mongodb
uv add langgraph-store-mongodb

from langgraph.checkpoint.mongodb import MongoDBSaver
from langgraph.store.mongodb import MongoDBStore
from pymongo import MongoClient

# 写法一：
mongouri = f'mongodb://{username}:{password}@127.0.0.1:27017/news?authSource=admin'
with (
    MongoDBSaver.from_conn_string(mongouri, db_name="news", collection_name="checkpointer") as checkpointer,
    MongoDBStore.from_conn_string(mongouri, db_name="news", collection_name="store", ) as store,  # 长期记忆数据存放的集合名
):
    # 编译图：挂接检查点 + 长期存储
    app = builder.compile(checkpointer=checkpointer, store=store)

    # config 中的 thread_id 是检查点的唯一标识（必填）
    # 同一个 thread_id 会恢复之前的状态；不同 thread_id 相互隔离
    config = {"configurable": {"thread_id": "thread-001"}, "user_id": "68"}
    # invoke: 一次性执行完整图
    result = app.invoke({"messages": ["hello"], "count": 0}, config=config)

# 写法二
mongouri = f'mongodb://{username}:{password}@127.0.0.1:27017/graph?authSource=admin'
mongo_client = MongoClient(mongouri)
collection = mongo_client["graph"]["checkpoints"]   # 库名["graph"] + 集合名["checkpoints"]
collection2 = mongo_client["graph"]["store"]  
checkpointer = MongoDBSaver(collection)
store = MongoDBStore(collection2)

## 查询记忆
for memory in store.search(("memories", "68")):  # namespace
    print(333, memory.value)
# 333 {'data': '您好，我是刘明'}
# 333 {'data': '您好，刘明！很高兴认识您。请问有什么可以帮您？'}

# 查看历史（检查点记录）
for checkpoint in app.get_state_history(config):
    print(563, checkpoint)
# StateSnapshot(values={'messages': ['hello', '来自 node_a，第 0 次调用', '来自 node_b'], 'count': 1}, next=(), config={'configurable': {'thread_id': 'thread-001', 'checkpoint_ns': '', 'checkpoint_id': '1f1c7820-73a7-652d-8002-5b2891609753'}}, metadata={'source': 'loop', 'step': 2, 'parents': {}}, created_at='2026-10-14T03:47:50.614300+00:00', parent_config={'configurable': {'thread_id': 'thread-001', 'checkpoint_ns': '', 'checkpoint_id': '1f1c7820-739b-6022-8001-3bf17713d514'}}, tasks=(), interrupts=())