Checkpointing

Checkpoints allow LangGraph agents to persist their state within and across multiple interactions. A checkpoint is a snapshot of the graph state at a given point in time, identified by a unique, monotonically increasing ID.

Learn more: For conceptual guides and tutorials on persistence, see the Persistence documentation.

Core Concepts

Concept	Description
Checkpoint	Snapshot of graph state at a point in time, including channel values, channel versions, and version tracking per node
Thread	A sequence of checkpoints identified by a unique thread_id; optionally narrow to a specific checkpoint with checkpoint_id
Serde	Serialization protocol used to encode/decode checkpoint data; defaults to JsonPlusSerializer
Pending writes	Writes from successful nodes preserved when other nodes fail, enabling resumption without re-running completed work

Checkpoint Savers

LangGraph provides several checkpoint saver implementations:

Saver	Package	Async Support	Use Case
InMemorySaver	langgraph-checkpoint	Yes	Debugging and testing only
SqliteSaver	langgraph-checkpoint-sqlite	No	Lightweight demos and small projects
AsyncSqliteSaver	langgraph-checkpoint-sqlite	Yes	Async SQLite; not recommended for production
PostgresSaver	langgraph-checkpoint-postgres	No	Production workloads with full checkpoint history
AsyncPostgresSaver	langgraph-checkpoint-postgres	Yes	Async production workloads with full checkpoint history

Quick Examples

In-Memory (Testing)

from langgraph.checkpoint.memory import MemorySaver  checkpointer = MemorySaver() graph = builder.compile(checkpointer=checkpointer)

PostgreSQL (Production)

from langgraph.checkpoint.postgres import PostgresSaver  with PostgresSaver.from_conn_string("postgresql://user:pass@localhost/db") as checkpointer:  checkpointer.setup()  graph = builder.compile(checkpointer=checkpointer)

Async PostgreSQL

from langgraph.checkpoint.postgres.aio import AsyncPostgresSaver  async with AsyncPostgresSaver.from_conn_string("postgresql://user:pass@localhost/db") as checkpointer:  await checkpointer.setup()  graph = builder.compile(checkpointer=checkpointer)

Serialization

All checkpoint savers use a serialization protocol to encode and decode checkpoint data. The default serializer is JsonPlusSerializer, which uses ormsgpack with a fallback to an extended JSON format that handles LangChain/LangGraph types, datetimes, enums, and more.

For sensitive data, wrap any serializer with EncryptedSerializer to encrypt checkpoint contents at rest:

from langgraph.checkpoint.serde.encrypted import EncryptedSerializer from langgraph.checkpoint.postgres import PostgresSaver  encrypted_serde = EncryptedSerializer.from_pycryptodome_aes(encryption_key) checkpointer = PostgresSaver(conn, serde=encrypted_serde)

Custom Checkpoint Savers

Subclass BaseCheckpointSaver and implement the following methods:

• get_tuple — Retrieve a specific checkpoint
• list — List checkpoints, optionally filtered
• put — Store a checkpoint
• put_writes — Store pending writes
• delete_thread — Delete all checkpoints for a thread

For async usage, implement the corresponding async variants (aget_tuple, alist, aput, aput_writes, adelete_thread).

API Reference