AI and LLMAgents

LangGraph Checkpoints and Long-Term Memory

Persist LangGraph agents on HotKV. Pause, resume and time-travel through graph state, and keep long-term memory searchable by meaning.

conformance tests passed: 81
savers: Sync + async
needed: No license

Step 1 of 6

Install

One package for the checkpointer and the store. Python 3.10 or newer.

Example sessionplaying

# Install

$ pip install langgraph-checkpoint-hotkv

Successfully installed langgraph-checkpoint-hotkv

An animated walkthrough with real HotKV commands and example replies. Click any step to jump to it.

The problem

A LangGraph agent that keeps its state in process memory loses everything on a restart. Human-in-the-loop approvals cannot wait overnight, failed runs cannot be resumed, and nothing the agent learned survives to the next conversation.

How HotKV handles it

langgraph-checkpoint-hotkv stores each graph's checkpoints in HotKV. A graph can be interrupted and resumed, or rewound to any earlier step and branched from it. The companion store keeps long-term memory across threads, found by namespace, by filter or, with an embedding model, by meaning.

Step by step

What happens, command by command

01
Install
One package for the checkpointer and the store. Python 3.10 or newer.
pip install langgraph-checkpoint-hotkv
02
Compile with a HotKV checkpointer
Every step of the graph is now saved in HotKV.
graph = builder.compile(checkpointer=HotKVSaver(hotkv.from_url("hotkv://localhost:6379")), interrupt_before=["refund"])
03
Run until it needs approval
The graph works through its steps, saving a checkpoint after each, and stops before the refund.
graph.invoke({"messages": [("user", "Refund order 1182")]}, {"configurable": {"thread_id": "t-81"}})
04
Hours later, after a restart
The process restarted overnight. The state is still in HotKV, waiting at the same step.
graph.get_state({"configurable": {"thread_id": "t-81"}}).next
05
Approve and resume
The approver says yes and the graph continues from the saved checkpoint.
graph.invoke(None, {"configurable": {"thread_id": "t-81"}})
06
Remember across threads
Long-term memory lives in the store and is found again in any later conversation.
store.put(("users", "42"), "prefs", {"refunds": "prefers store credit"})

Where it fits

Real-world scenarios

Human-in-the-loop workflows

Refunds, contract changes and production deployments wait for a person to approve, for minutes or days, without holding a process open.

Long-running research agents

A run that fails at step 40 resumes from step 40 instead of starting over.

Debugging and evaluation

Go back to any earlier checkpoint, change the input and branch, to see how the agent would have behaved.

Built in

What you get

Checkpointer

HotKVSaver and AsyncHotKVSaver keep graph state between runs, with interrupt, resume and time travel.

Long-term store

HotKVStore keeps items across threads, searchable by namespace, filter or meaning, with optional expiry.

Conformance tested

Passes LangGraph's own checkpointer conformance suite, 81 tests in 8 capabilities, for both savers.

No modules, no license

Uses HotKV's core data types and its vector index, and works with or without an enterprise license.

Checkpointer and store

python

from langgraph.checkpoint.hotkv import HotKVSaver

with HotKVSaver.from_conn_string("hotkv://localhost:6379") as checkpointer:
    graph = builder.compile(checkpointer=checkpointer)
    graph.invoke({"messages": [("user", "hi")]}, {"configurable": {"thread_id": "1"}})

Reference:hotkv-langgraph on GitHub