In-Memory-Checkpointer.md

In-Memory Checkpointer

BoundedInMemorySaver, also exported as InMemorySaver, extends LangGraph's in-memory saver with resource and concurrency guardrails. Its storage lifetime matches the current process.

Defaults

Control	Default	Behavior
Idle thread TTL	4 hours	Removes a complete thread during a later saver operation or purge_expired()
Thread limit	1,000	Evicts the least-recently-used complete thread
Namespaces per thread	16	Rejects an additional namespace
Checkpoints per thread	1,000	Rejects an additional checkpoint
Checkpoint payload	1 MiB	Rejects before storage mutation
Pending-write batch	512 KiB	Rejects before storage mutation
Total serialized payload	64 MiB	Evicts other LRU threads, then rejects if necessary
Identifier length	256 characters	Rejects invalid or oversized identifiers

The saver uses a re-entrant lock around shared storage. Async methods move sync serialization and storage work to a worker thread. It uses LangGraph's JsonPlusSerializer with pickle_fallback=False.

Eviction operates at complete-thread granularity to preserve checkpoint ancestry and channel versions. Capacity failures raise CheckpointCapacityError while leaving the protected thread unchanged.

Byte limits measure serialized payloads. Monitor process resident memory alongside the saver statistics for production capacity planning.

Configuration

from fastapi_langgraph_server import (
    InMemoryCheckpointConfig,
    InMemorySaver,
    LangGraphServerConfig,
)

checkpointer = InMemorySaver(
    config=InMemoryCheckpointConfig(
        idle_ttl_seconds=30 * 60,
        max_threads=250,
        max_checkpoints_per_thread=500,
        max_checkpoint_bytes=512 * 1024,
        max_pending_write_bytes=256 * 1024,
        max_total_bytes=32 * 1024 * 1024,
    )
)

config = LangGraphServerConfig(
    assistants={assistant.assistant_id: assistant},
    checkpointer=checkpointer,
)

stats = checkpointer.stats()
expired = checkpointer.purge_expired()

Instantiate and pass the saver to enable process-local persistence. Server configurations use stateless execution when checkpointer is omitted.

Suitable workloads

The in-memory saver fits:

• local development and integration tests;
• single-process applications;
• ephemeral workflows whose state lifetime matches the process; and
• bounded workloads that benefit from TTL and LRU eviction.

For shared workers, recovery, or durable state, configure a compatible persistent async saver. Authenticate callers, authorize thread ownership, limit request sizes and rates, keep sensitive values out of state, and monitor capacity. create_app supplies request-body and run-concurrency limits; configure an execution timeout and per-client rate limit for the workload.