LangChain 与 Mem0 集成长期记忆

1uv add mem0ai chromadb ollama

 1import os
 2import logging
 3
 4os.environ["MEM0_TELEMETRY"] = "false"
 5logging.getLogger("mem0.utils.spacy_models").setLevel(logging.ERROR)
 6
 7from mem0 import Memory
 8
 9config = {
10    "vector_store": {
11        "provider": "chroma",
12        "config": {
13            "collection_name": "mem0_memories",
14            "path": "./chroma_mem0",
15        },
16    },
17    "llm": {
18        "provider": "ollama",
19        "config": {
20            "model": "llama3.1:8b",
21        },
22    },
23    "embedder": {
24        "provider": "ollama",
25        "config": {
26            "model": "embeddinggemma:latest",
27        },
28    },
29}
30
31client = Memory.from_config(config)
32
33# Add a memory
34messages = [
35    {"role": "user", "content": "我不是一个素食主义者，但我喜欢吃蔬菜。"},
36    {"role": "assistant", "content": "Got it! I'll remember your dietary preferences."},
37]
38client.add(messages, user_id="user123")
39
40# Search memories
41results = client.search("What are my dietary restrictions?", filters={"user_id": "user123"})
42print(results)

1tree chroma_mem0
2chroma_mem0
3├── 6227fa08-f436-40e5-b474-bb78ec2012ce
4│   ├── data_level0.bin
5│   ├── header.bin
6│   ├── length.bin
7│   └── link_lists.bin
8└── chroma.sqlite3

 1sqlite3 chroma_mem0/chroma.sqlite3
 2SQLite version 3.51.0 2025-06-12 13:14:41
 3Enter ".help" for usage hints.
 4sqlite> .table
 5acquire_write                      embedding_metadata_array
 6collection_metadata                embeddings
 7collections                        embeddings_queue
 8databases                          embeddings_queue_config
 9embedding_fulltext_search          maintenance_log
10embedding_fulltext_search_config   max_seq_id
11embedding_fulltext_search_content  migrations
12embedding_fulltext_search_data     segment_metadata
13embedding_fulltext_search_docsize  segments
14embedding_fulltext_search_idx      tenants
15embedding_metadata
16sqlite> .header on
17sqlite> select * from embedding_metadata;
18id|key|string_value|int_value|float_value|bool_value
191|updated_at|2026-05-02T16:17:05.745485+00:00|||
201|created_at|2026-05-02T16:17:05.745485+00:00|||
211|hash|1cfd6173c20814e47af50c62ab626ac6|||
221|user_id|user123|||
231|attributed_to|user|||
241|data|用户指出自己不是素食主义者，但喜欢吃蔬菜|||
251|text_lemmatized|用户指出自己不是素食主义者，但喜欢吃蔬菜|||

 1from typing import Any
 2
 3from langchain.agents import create_agent
 4from langchain.agents.middleware import AgentMiddleware, AgentState
 5from langchain_core.messages import SystemMessage, HumanMessage, AIMessage
 6from langgraph.runtime import Runtime
 7from mem0 import Memory
 8from pydantic import BaseModel
 9
10config = ... # config 与上相同
11
12class Context(BaseModel):
13    user_id: str
14
15memory = Memory.from_config(config)
16
17class Mem0Middleware(AgentMiddleware):
18
19    def before_model(self, state: AgentState, runtime: Runtime[Context]) -> dict[str, Any] | None:
20        user_id = runtime.context.user_id
21        last_message = state["messages"][-1]
22        last_user_msg = last_message.content if isinstance(last_message, HumanMessage) else ""
23        if not last_user_msg:
24            return None
25
26        results = memory.search(query=last_user_msg, filters={"user_id": user_id}, limit=5)
27        if not results["results"]:
28            return None
29
30        mem_str = "\n".join([f"- {m['memory']}" for m in results["results"]])
31        memory_msg = SystemMessage(
32            content=f"[user long-term memory]\n{mem_str}",
33            id="mem0_context",
34        )
35
36        return {"messages": [memory_msg, *state["messages"]]}
37
38    def after_model(self, state: AgentState, runtime: Runtime[Context]) -> dict[str, Any] | None:
39        user_id = runtime.context.user_id
40        msgs = state["messages"]
41        recent = [
42            {"role": "user" if m.type == "human" else "assistant",
43             "content": m.content}
44            for m in msgs[-2:] if isinstance(m, (HumanMessage, AIMessage)) and m.content
45        ]
46        if recent:
47            result = memory.add(recent, user_id=user_id)
48            print(result)
49
50        return None
51
52
53agent = create_agent(
54    model="ollama:gemma4:e4b",
55    middleware=[Mem0Middleware()],
56    context_schema=Context,
57    # store=memory,
58    system_prompt="You are an assistant with long-term memory, please answer question as concise as you can.",
59)
60
61result = agent.invoke(
62    {"messages": [{"role": "user", "content": "I'm not a vegetarian, but allergic to nuts."}]},
63    context=Context(user_id="user123")
64)
65print("-" * 50)
66print(result["messages"][-1].content)
67
68result = agent.invoke(
69    {"messages": [{"role": "user", "content": "What are my dietary restrictions?"}]},
70    context=Context(user_id="user123")
71)
72print("-" * 50)
73print(result["messages"][-1].content)

1memory.add(recent, user_id=user_id)

 1@tool
 2def save_memory(content: str, user_id: str) -> str:
 3    """Save a long-term fact about the user to the memory store.
 4    Call this when the user reveals preferences, attributes, or important events."""
 5    result = memory.add(messages=[{"role": "user", "content": content}], user_id=user_id)
 6    print(result)
 7    return f"Saved: {content}"
 8
 9@tool
10def search_memory(query: str, user_id: str) -> str:
11    """Search the memory store for user history relevant to the query.
12    Call this before answering personalized questions."""
13    results = memory.search(query=query, filters={"user_id": user_id}, limit=5)
14    if not results["results"]:
15        return "No relevant memories found."
16    return "\n".join([f"- {m['memory']}" for m in results["results"]])
17
18agent = create_agent(
19    model="ollama:gemma4:e4b",
20    tools=[save_memory, search_memory],
21    context_schema=Context,
22    system_prompt=(
23        "You are an assistant with long-term memory. "
24        "Before answering personalized questions, call search_memory first. "
25        "When the user shares important information, call save_memory to persist it. "
26        "Current user ID: {user_id}"
27    ),
28)