LangGraph

Give a LangGraph agent long-term memory across threads. Ships in the Python SDK.

Install

pip install "suprflo[langgraph]"

Targets langgraph >= 0.2.

What you get

All three are built on the Python SDK MemoryClient:

• SuprfloStore — a BaseStore implementation, so get / put / search / delete inside a graph read and write Suprflo memories. Read the limits below before using it.
• create_memory_node(client, user_id) — a graph node that recalls memories relevant to the last message into state["memories"].
• create_memory_writer_node(client, user_id) — a graph node that persists state["messages"] to Suprflo.

Store

from suprflo import MemoryClient
from suprflo.integrations.langgraph import SuprfloStore

client = MemoryClient(api_key="YOUR_API_KEY")
store = SuprfloStore(client=client)

# Namespaces map onto Suprflo subjects: the last element is the user_id.
store.search(("memories", "alice"), query="hobbies")

SuprfloStore is a memory store, not a faithful key-value backend. Three gaps to know about:

• put cannot create by key. Suprflo assigns memory ids server side, so a put with an unknown key raises KeyError (it cannot create a memory at a key you choose). put on an existing key updates it in place, which does behave as expected. Take keys from search results rather than choosing them yourself.
• value is flattened to text. Suprflo stores a memory string, not an arbitrary JSON blob. A value's "data" field (or the whole value stringified, if absent) is what gets persisted; other fields become metadata on create and are dropped on update.
• list_namespaces raises NotImplementedError — Suprflo has no namespace tree to enumerate.

abatch calls the synchronous HTTP client, so it blocks the event loop. Fine for typical graph fan-out; not for high concurrency.

Nodes in a graph

from langgraph.graph import END, START, StateGraph
from suprflo.integrations.langgraph import create_memory_node, create_memory_writer_node

builder = StateGraph(State)
builder.add_node("recall", create_memory_node(client, user_id="alice", top_k=5))
builder.add_node("agent", my_agent_node)          # reads state["memories"]
builder.add_node("save", create_memory_writer_node(client, user_id="alice"))

builder.add_edge(START, "recall")
builder.add_edge("recall", "agent")
builder.add_edge("agent", "save")
builder.add_edge("save", END)

recall returns {"memories": [...]} (a list of memory strings) for a downstream prompt to consume; save returns {} and leaves graph state untouched. Both read plain dict messages and LangChain message objects.

If langgraph isn't installed, importing this module raises a clear ImportError telling you to pip install "suprflo[langgraph]".