LangGraph Checkpointing (Python)

Overview

A chat agent that "keeps introducing itself" is almost always P16. The caller invokes graph.invoke(state) without passing config={"configurable": {"thread_id": ...}} — LangGraph's checkpointer silently spawns a fresh state per call. No error, no warning, no log line. The user sees it; the code does not.

That is one of five separate checkpointing pitfalls this skill covers:

• P16 — missing thread_id silently resets memory
• P17 — interrupt_before raises TypeError when state holds non-JSON values (datetime, Decimal, custom classes) — and it raises at the interrupt boundary, not when the bad value was first assigned, so the traceback points at the wrong line
• P20 — PostgresSaver does not auto-migrate checkpoint schema; upgrading langgraph silently reads old checkpoints as empty state
• P40 — ConversationBufferMemory and the rest of legacy chat memory were removed in LangChain 1.0; checkpointers are the replacement
• P51 — Deep Agent virtual-FS state in state["files"] grows unboundedly and eventually makes checkpoint writes a latency hotspot

This skill walks through picking a checkpointer by environment, enforcing thread_id at the application boundary, constraining state to JSON-safe primitives, Postgres setup + migration, and time-travel for incident debugging. Pinned to langgraph >= 1.0, < 2.0, langgraph-checkpoint-postgres >= 1.0, < 2.0. Pain-catalog anchors: P16, P17, P18, P20, P22, P40, P51.

Prerequisites

• Python 3.10+
• pip install langgraph langchain-core (both >= 1.0, < 2.0)
• For Postgres: pip install langgraph-checkpoint-postgres and a Postgres 13+ instance
• For async Postgres: the same package plus asyncpg
• A thread_id strategy — typically a UUID4 string per conversation; see thread-id-discipline.md

Instructions

Step 1 — Pick a checkpointer by environment

MemorySaver is in-process only. State vanishes on restart. Every worker has its own (P22 analog for LangGraph). Use it anywhere state loss is acceptable; never in a multi-worker web backend.

PostgresSaver and its async sibling require setup() on every startup and after every langgraph upgrade (see Step 5). Checkpoint storage overhead is typically 1-10 KB per step of serialized state; plan your DB size accordingly — a 2,000-turn conversation with 3 KB average state fits in ~6 MB per thread.

See checkpointer-comparison.md for the full matrix including latency, concurrency, and the FastAPI lifespan pattern.

Step 2 — Require thread_id at every invocation

This is the fail-loud middleware that prevents P16:

from typing import Any

def require_thread_id(config: dict[str, Any]) -> dict[str, Any]:
    """Raise if thread_id is missing. Fails loud so P16 surfaces in tests,
    not in user-visible conversation logs."""
    configurable = (config or {}).get("configurable", {})
    thread_id = configurable.get("thread_id")
    if not thread_id:
        raise ValueError(
            "thread_id missing from config['configurable']. "
            "Every graph invocation must carry a thread_id."
        )
    if not isinstance(thread_id, str):
        raise TypeError(
            f"thread_id must be str (UUID), got {type(thread_id).__name__}"
        )
    return config

Call it at every application boundary:

import uuid

config = {"configurable": {"thread_id": str(uuid.uuid4())}}
require_thread_id(config)
result = graph.invoke(initial_state, config=config)

For web endpoints, extract to a FastAPI dependency (Header(...) with no default — forces 422 on missing). For multi-tenant apps, scope the thread id by composing tenant + user + conversation ids into a single colon-delimited string (example: "acme:alice:conv-1"). See thread-id-discipline.md for UUID generation, rotation, and the integration test that proves tenants do not share state.

Step 3 — Keep state JSON-serializable (TypedDict, primitives only)

from typing import Annotated, TypedDict
from langgraph.graph.message import add_messages
from langchain_core.messages import AnyMessage


class AgentState(TypedDict):
    # Messages are safe — LangGraph registers a custom serializer.
    messages: Annotated[list[AnyMessage], add_messages]

    # Primitives only below. NO datetime, NO Decimal, NO custom classes.
    user_id: str
    turn_count: int
    last_action_at: str           # ISO string, not datetime.datetime
    pending_approval: bool
    metadata: dict[str, str]      # dict keys must be str
    plan: list[dict[str, str]]    # list of primitive dicts

The rule: state fields must be JSON-safe primitives or recursive structures of them (str, int, float, bool, None, list, dict[str, ...]). json.dumps(state) must succeed. If it raises, the checkpointer raises — often at a HITL interrupt many steps later (P17), which is why the traceback never points at the line that introduced the bad value.

For non-primitive inputs, coerce at node output boundaries with a helper:

from datetime import datetime
from decimal import Decimal

def record_purchase(state: AgentState) -> dict:
    now = datetime.utcnow()
    price = Decimal("19.99")
    return {
        "last_action_at": now.isoformat(),
        "metadata": {**state["metadata"], "price": str(price)},
    }

Forbidden-types reference and the full to_state / from_state helper pair are in json-serializability-rules.md.

Step 4 — Compile the graph with a checkpointer (Postgres, sync)

from langgraph.checkpoint.postgres import PostgresSaver
from langgraph.graph import StateGraph
import os

DB_URI = os.environ["DATABASE_URL"]

def build_graph() -> StateGraph:
    builder = StateGraph(AgentState)
    builder.add_node("agent", agent_node)
    builder.add_node("human_approval", human_approval_node)
    builder.set_entry_point("agent")
    builder.add_edge("agent", "human_approval")
    builder.set_finish_point("human_approval")
    return builder

with PostgresSaver.from_conn_string(DB_URI) as checkpointer:
    checkpointer.setup()     # Idempotent. Creates checkpoint tables if missing.
    graph = build_graph().compile(
        checkpointer=checkpointer,
        interrupt_before=["human_approval"],
    )

    config = {"configurable": {"thread_id": "user-123"}}
    require_thread_id(config)
    result = graph.invoke({"messages": [HumanMessage("hi")]}, config=config)

For async, mirror the pattern with AsyncPostgresSaver.from_conn_string(...) inside a FastAPI @asynccontextmanager lifespan; every call site uses await graph.ainvoke(...).

Step 5 — Run setup() on startup AND after every langgraph upgrade

P20 is the quiet one: you pip install --upgrade langgraph, tests pass, CI goes green, you deploy. Existing threads come back empty. No DB error.

# Put this in your deploy script / migration runbook:
with PostgresSaver.from_conn_string(DB_URI) as checkpointer:
    checkpointer.setup()
    # Sanity check: read one known thread and assert it's not empty.
    snap = checkpointer.get({"configurable": {"thread_id": "canary-thread"}})
    assert snap is not None, "Canary thread lost after schema migration"

Run this in staging first, with a canary thread whose state you pre-populated from an older langgraph version. If the assertion holds, promote. If not, the migration path is: dump checkpoints, upgrade, restore via a migration script. Do not promote to production on a version bump without this check.

Step 6 — Time-travel for incident debugging

Every checkpoint is keyed by (thread_id, checkpoint_id) and reachable via graph.get_state_history(config):

config = {"configurable": {"thread_id": bad_thread_id}}
history = list(graph.get_state_history(config))

for i, snap in enumerate(history):
    print(i, snap.metadata.get("step"), snap.next, list(snap.values.keys()))

# Resume from a specific prior checkpoint — None as input = "use this state":
past = history[3]                                    # history is newest-first
result = graph.invoke(None, config=past.config)       # past.config pins checkpoint_id

To fix state and replay:

graph.update_state(
    past.config,
    {"retry_count": 0, "error_reason": None},
    as_node="validator",
)
graph.invoke(None, config=past.config)

The original branch is preserved — update_state creates a new checkpoint alongside the old one. Useful for forensics and for A/B comparing an old vs new model on the exact same state.

See time-travel-and-replay.md for the full incident playbook, branching for A/B eval, and checkpoint pruning SQL.

Output

• A checkpointer selected by environment, not copy-pasted from a tutorial
• A require_thread_id middleware that fails loud on missing thread_id, and at least one integration test that asserts two tenants do not share state
• An AgentState TypedDict constrained to JSON-safe primitives, plus a to_state helper at node output boundaries for datetime / Decimal / Pydantic / enums
• PostgresSaver.setup() wired into startup and every deploy that bumps langgraph
• An incident runbook that uses get_state_history + update_state to time-travel an agent state, with prior branches preserved

Error Handling

Examples

Minimal chat agent with MemorySaver

Dev-mode multi-turn chat that survives within a single process. Good for local iteration and pytest fixtures; state vanishes on restart.

from langgraph.checkpoint.memory import MemorySaver
from langgraph.prebuilt import create_react_agent
from langchain_anthropic import ChatAnthropic

checkpointer = MemorySaver()
agent = create_react_agent(
    model=ChatAnthropic(model="claude-sonnet-4-6"),
    tools=[...],
    checkpointer=checkpointer,
)
config = {"configurable": {"thread_id": "dev-session-1"}}
agent.invoke({"messages": [HumanMessage("hi")]}, config=config)
agent.invoke({"messages": [HumanMessage("remember my name is alice")]}, config=config)
agent.invoke({"messages": [HumanMessage("what's my name?")]}, config=config)  # alice

Multi-tenant Postgres + FastAPI lifespan

Async production shape. One AsyncPostgresSaver pool, thread-id extracted from a required header, tenants isolated by a composite key.

See checkpointer-comparison.md for the complete FastAPI lifespan pattern and pool sizing advice (max_size needs to exceed concurrent graph count, and the LangGraph pool should be separate from the application's primary DB pool).

Migrating from ConversationBufferMemory (P40)

Legacy 0.x pattern replaced end-to-end:

# OLD — raises ImportError on 1.0:
# from langchain.memory import ConversationBufferMemory
# memory = ConversationBufferMemory()
# chain = LLMChain(llm=llm, memory=memory)

# NEW:
from langgraph.prebuilt import create_react_agent
from langgraph.checkpoint.postgres import PostgresSaver

with PostgresSaver.from_conn_string(DB_URI) as cp:
    cp.setup()
    agent = create_react_agent(model=llm, tools=tools, checkpointer=cp)
    config = {"configurable": {"thread_id": user_session_id}}
    agent.invoke({"messages": [HumanMessage(user_input)]}, config=config)

The thread_id is now the "session key" that ConversationBufferMemory used to carry implicitly.

Incident time-travel

Production returned a wrong answer for thread_id=bad-thread at 14:03. Walk history newest-first, inspect the prior state, patch and replay:

config = {"configurable": {"thread_id": "bad-thread"}}
history = list(graph.get_state_history(config))
for i, snap in enumerate(history):
    print(i, snap.metadata.get("step"), snap.next)
# Assume step 7 wrote the bad output; step 6 is the input.
prior = history[-7]  # or index by metadata["step"]
print("input to bad node:", prior.values)
graph.update_state(prior.config, {"retry_count": 0}, as_node="validator")
graph.invoke(None, config=prior.config)  # new branch with correct result

Full playbook (finding thread_id in logs, inspecting writes metadata, pruning history) in time-travel-and-replay.md.

Resources

• LangGraph persistence concepts
• Checkpointer API reference
• LangGraph HITL concepts
• LangGraph 1.0 release notes
• Pack pain catalog: docs/pain-catalog.md (entries P16, P17, P18, P20, P22, P40, P51)
• Related skills in this pack: langchain-langgraph-basics, langchain-langgraph-agents, langchain-upgrade-migration, langchain-common-errors