An agent that remembers across conversations

Give an agent the memory tools, a store and a checkpointer. A preference saved in one conversation thread can be searched from a new one.

Example

from langchain.agents import create_agent
from langgraph.checkpoint.memory import InMemorySaver
from langmem import create_manage_memory_tool, create_search_memory_tool

from memory_model import MemoryModel

namespace = ("memories", "{user_id}")
agent = create_agent(
    MemoryModel(),
    tools=[create_manage_memory_tool(namespace=namespace), create_search_memory_tool(namespace=namespace)],
    store=store,
    checkpointer=InMemorySaver(),
)

create_agent from LangChain runs the model in a loop with its tools. The tools were created without a store: inside the agent they use the store passed here. InMemorySaver is the checkpointer, keeping each thread's messages.

Example

first = {"configurable": {"user_id": "asha", "thread_id": "chat-1"}}
result = agent.invoke({"messages": [{"role": "user", "content": "Please email me about refunds."}]}, config=first)
for message in result["messages"]:
    print(f"{message.type:5} {message.content or message.tool_calls[0]['name']}"[:70])

The model called manage_memory, the tool saved the memory, and the model answered. That was thread chat-1.

Example

second = {"configurable": {"user_id": "asha", "thread_id": "chat-2"}}
result = agent.invoke({"messages": [{"role": "user", "content": "How do I like to be contacted?"}]}, config=second)
print(len(result["messages"]))
print(result["messages"][-1].content)

A new thread starts with no messages from the first: four messages here, the question, the search call, its result and the answer. The answer came from search_memory, which read Asha's namespace in the shared store. That is the difference between the checkpointer, per thread, and the store, per user.

Example

other = {"configurable": {"user_id": "ravi", "thread_id": "chat-3"}}
result = agent.invoke({"messages": [{"role": "user", "content": "How do I like to be contacted?"}]}, config=other)
print(result["messages"][-1].content)

Ravi's search returned an empty list, and the answer says so. The memory tools searched ("memories", "ravi"), so Asha's preference could not be found from Ravi's conversation.

Try it yourself

Ask Asha's question again in chat-1 and count the messages.
Add a system_prompt telling the agent to save contact preferences.
Print store.search(("memories", "asha")) after the first call.

PreviousMemory tools: manage and search

Slow is fine. Stopping is the only problem.