Agent memory

Agent memory is long-term memory an agent manages itself: LangMem's memory tools plus a store make a preference saved in one conversation thread findable from any later thread for the same user.

Last updated: 30 Sep, 2026 · LangMem 0.0.30

The memory tools become useful inside an agent. LangChain's create_agent runs the model in a loop with tools; with a checkpointer it keeps each thread's messages, and with a store it gives LangMem's tools somewhere to save. A new thread starts with no messages, so anything the agent knows about the customer there came from the store.

Syntax:

agent = create_agent(model, tools=[manage_tool, search_tool], store=store, checkpointer=InMemorySaver())
agent.invoke({"messages": [...]}, config={"configurable": {"user_id": "asha", "thread_id": "chat-1"}})

The agent

The tools are created without a store; inside the agent they use the one passed to create_agent. The system prompt says what to save and keeps answers to what the tools returned.

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

A config per thread

user_id picks the namespace; thread_id picks the conversation.

first = {"configurable": {"user_id": "ravi", "thread_id": "chat-1"}}
second = {"configurable": {"user_id": "ravi", "thread_id": "chat-2"}}

Saving in one thread, recalling in another

from langchain.agents import create_agent
from langchain.chat_models import init_chat_model
from langgraph.checkpoint.memory import InMemorySaver
from langgraph.store.memory import InMemoryStore
from langmem import create_manage_memory_tool, create_search_memory_tool

PROMPT = "You are a support assistant for an online shop. When the customer tells you a preference, save it with manage_memory as one plain sentence. Before answering a question about the customer, search memory. Answer in one short sentence, using only what the tools returned."

model = init_chat_model("groq:openai/gpt-oss-120b", temperature=0)
store = InMemoryStore(index={"dims": 3072, "embed": "google_genai:gemini-embedding-2"})
namespace = ("memories", "{user_id}")
agent = create_agent(
    model,
    tools=[create_manage_memory_tool(namespace=namespace), create_search_memory_tool(namespace=namespace)],
    store=store,
    checkpointer=InMemorySaver(),
    system_prompt=PROMPT,
)


def chat(user_id, thread_id, text):
    config = {"configurable": {"user_id": user_id, "thread_id": thread_id}}
    result = agent.invoke({"messages": [{"role": "user", "content": text}]}, config=config)
    print(f"[{user_id} / {thread_id}] {text}")
    for message in result["messages"][1:]:
        what = message.content or ", ".join(call["name"] for call in message.tool_calls)
        print(f"  {message.type:5} {what[:90]}")


chat("ravi", "chat-1", "Please call me about refunds, not email.")
chat("ravi", "chat-2", "How do I like to be contacted?")
chat("meera", "chat-3", "How do I like to be contacted?")

[ravi / chat-1] Please call me about refunds, not email.
  ai    manage_memory
  tool  created memory 4fabe5b0-4b3b-46b3-80f5-2fd61240cc6b
  ai    Got it—I'll call you about refunds instead of emailing.
[ravi / chat-2] How do I like to be contacted?
  ai    search_memory
  tool  [{"namespace":["memories","ravi"],"key":"4fabe5b0-4b3b-46b3-80f5-2fd61240cc6b","value":{"c
  ai    You prefer to be called about refunds, not emailed.
[meera / chat-3] How do I like to be contacted?
  ai    search_memory
  tool  []
  ai    I have no record of your contact preference.

What happened in each thread

• chat-1: the model called manage_memory, the tool created a memory in Ravi's namespace, and the reply confirmed it.
• chat-2, a new thread: the checkpointer gave the agent none of chat-1's messages. It called search_memory, the tool returned the saved memory, and the answer came from it: call about refunds, not email.
• chat-3, Meera: the same question searched Meera's namespace, the tool returned [], and the agent said it had no record instead of guessing, because the prompt limits answers to what the tools returned.

Checkpointer vs store in an agent

When an agent should manage its own memory

• A personal assistant that users talk to over weeks, in separate chats.
• Support, where the customer's preferences should follow them to the next ticket.
• Any agent whose users say "remember that...".

Related

• Previous: Memory tools
• Next: Background memory
• Reference: Hot path quickstart