Memory tools

Memory tools are two LangChain tools from LangMem, manage_memory and search_memory, that let an agent save, update, delete and search memories in the store itself, during the conversation.

Last updated: 30 Sep, 2026 · LangMem 0.0.30

A store manager decides after a chat what to remember. Memory tools move that decision into the conversation: the agent calls manage_memory when the customer says something worth keeping, and search_memory before answering. This lesson calls the tools directly, the way an agent would, to see what each one takes and returns.

Syntax:

save = create_manage_memory_tool(namespace=("memories", "{user_id}"), store=store)
find = create_search_memory_tool(namespace=("memories", "{user_id}"), store=store)

The manage_memory tool

A tool is what a model sees: a name, a description and arguments. manage_memory takes content, an action of create, update or delete, and an id for the last two.

save = create_manage_memory_tool(namespace=("memories", "{user_id}"), store=store)
print(save.name, list(save.args))

Calling a tool with a config

The namespace template is filled from the config, as with the store manager.

asha = {"configurable": {"user_id": "asha"}}
save.invoke({"content": "Order A-1001 arrived broken", "action": "create"}, config=asha)

Saving and searching with the tools

import json

from langgraph.store.memory import InMemoryStore
from langmem import create_manage_memory_tool, create_search_memory_tool

store = InMemoryStore(index={"dims": 3072, "embed": "google_genai:gemini-embedding-2"})
save = create_manage_memory_tool(namespace=("memories", "{user_id}"), store=store)
find = create_search_memory_tool(namespace=("memories", "{user_id}"), store=store)
print(save.name, list(save.args), "|", find.name, list(find.args))

asha = {"configurable": {"user_id": "asha"}}
print(save.invoke({"content": "Order A-1001 arrived broken", "action": "create"}, config=asha)[:15])
print(save.invoke({"content": "Wants us to email her", "action": "create"}, config=asha)[:15])

results = json.loads(find.invoke({"query": "damaged parcel"}, config=asha))
for row in results:
    print(round(row["score"], 3), row["value"]["content"])

manage_memory ['content', 'action', 'id'] | search_memory ['query', 'limit', 'offset', 'filter']
created memory 
created memory 
0.652 Order A-1001 arrived broken
0.517 Wants us to email her

What the tools returned

• The tools' arguments are what an agent fills in: content, action and id for manage_memory; query, limit, offset and filter for search_memory.
• create returns text, "created memory" followed by the new id; the example prints only the first 15 characters, so the id is cut off. That text is what the agent reads back as the tool result.
• search returns JSON, best match first. "damaged parcel" found "Order A-1001 arrived broken" at 0.652, above the email preference at 0.517.

Memory tools vs a store manager

When to give an agent memory tools

• When users ask the assistant to remember or forget something explicitly.
• When the agent should look a fact up only if the question needs it.
• In multi-agent setups, where each agent writes to its own namespace and searches a shared one.

Related

• Previous: Store managers
• Next: Agent memory
• Reference: How to use memory tools