The LangGraph store: where memories live

LangGraph's store saves JSON documents under a namespace, a tuple like ("memories", "asha"), and a key. LangMem's stateful parts read and write it.

Example

from langgraph.store.memory import InMemoryStore

store = InMemoryStore()
store.put(("memories", "asha"), "contact", {"content": "Wants us to email them"})
store.put(("memories", "ravi"), "contact", {"content": "Prefers to be called"})

print(store.get(("memories", "asha"), "contact").value)
print([item.key for item in store.search(("memories", "asha"))])
print(store.list_namespaces())

put writes a dictionary under a namespace and key, get reads one back as an Item, and search lists the items in a namespace. Namespaces work like folders: Asha's memories and Ravi's never mix unless you search a shared prefix.

Example

store.put(("memories", "asha"), "contact", {"content": "Wants us to text them"})
print(store.get(("memories", "asha"), "contact").value)
store.delete(("memories", "asha"), "contact")
print(store.get(("memories", "asha"), "contact"))

A put to an existing key replaces the value, and delete removes it.

Stores for production

InMemoryStore is lost when the process ends. PostgresStore, from langgraph-checkpoint-postgres, has the same methods and keeps data in PostgreSQL; LangGraph Platform provides one automatically. Code written against the store works with either.

The store is not the checkpointer

LangGraph's checkpointer saves one conversation's state, a thread. The store is shared across threads, which is what makes memory long-term. Lesson 13 uses both.

Try it yourself

Search ("memories",) and see which items a prefix finds.
Store a nested dictionary and read one field back.
Print an item's created_at and updated_at after a second put.

PreviousSchemas and profiles

This is what real progress feels like.