langgraph-state-management（语言状态管理）

LangGraph State Management

State Design Workflow

Follow this workflow when designing or modifying state for a LangGraph application:

1. Identify data requirements — What data flows through the graph?
2. Choose a schema pattern — Match the use case to a template
3. Define reducers — Decide how concurrent updates merge
4. Configure persistence — Select and set up a checkpointer
5. Validate and test — Run schema validation and reducer tests

Quick Start

Python — Minimal Chat State

from langgraph.graph import StateGraph, START, END, MessagesState
from langchain_core.messages import AIMessage

class State(MessagesState):
    pass

def chat_node(state: State):
    return {"messages": [AIMessage(content="Hello!")]}

graph = StateGraph(State).add_node("chat", chat_node)
graph.add_edge(START, "chat").add_edge("chat", END)
app = graph.compile()

Python — Subclass MessagesState

For convenience, subclass the built-in MessagesState (includes messages with add_messages reducer):

from langgraph.graph import MessagesState

class State(MessagesState):
    documents: list[str]
    query: str

TypeScript — StateSchema with Zod

import { StateGraph, StateSchema, MessagesValue, ReducedValue, START, END } from "@langchain/langgraph";
import { AIMessage } from "@langchain/core/messages";
import { z } from "zod/v4";

const State = new StateSchema({
  messages: MessagesValue,
  documents: z.array(z.string()).default(() => []),
  count: new ReducedValue(
    z.number().default(0),
    { reducer: (current, update) => current + update }
  ),
});

const graph = new StateGraph(State)
  .addNode("chat", (state) => ({ messages: [new AIMessage("Hello!")] }))
  .addEdge(START, "chat")
  .addEdge("chat", END)
  .compile();

Schema Patterns

Choose the pattern matching the application type. See references/schema-patterns.md for complete examples with both Python and TypeScript.

Template files are available in assets/ for each pattern:

• assets/chat_state.py — Chat application
• assets/research_state.py — Research agent
• assets/workflow_state.py — Workflow orchestration
• assets/tool_calling_state.py — Tool-calling agent

For RAG state patterns, use reference examples in references/schema-patterns.md.

Reducers

Reducers control how state updates merge when nodes write to the same field.

Key Concepts

• No reducer → value is overwritten (last-write-wins)
• With reducer → values are merged using the reducer function
• A reducer takes (existing_value, new_value) and returns the merged result

Python: Annotated Type with Reducer

from typing import Annotated
import operator
from langgraph.graph import MessagesState

class State(MessagesState):
    # Overwrite (no reducer)
    query: str

    # Sum integers
    count: Annotated[int, operator.add]

    # Custom reducer
    results: Annotated[list[str], lambda left, right: left + right]

TypeScript: ReducedValue and MessagesValue

const State = new StateSchema({
  query: z.string(),                    // Last-write-wins
  messages: MessagesValue,              // Built-in message reducer
  count: new ReducedValue(              // Custom reducer
    z.number().default(0),
    { reducer: (current, update) => current + update }
  ),
});

Built-in Reducers

Bypass Reducers with Overwrite

Replace accumulated state instead of merging:

from langgraph.types import Overwrite

def reset_messages(state: State):
    return {"messages": Overwrite(["fresh start"])}

Delete Messages

from langchain_core.messages import RemoveMessage
from langgraph.graph.message import REMOVE_ALL_MESSAGES

# Delete specific message
{"messages": [RemoveMessage(id="msg_123")]}

# Delete all messages
{"messages": [RemoveMessage(id=REMOVE_ALL_MESSAGES)]}

For advanced reducer patterns (deduplication, deep merge, conditional update, size-limited accumulators), see references/reducers.md.

Persistence

Persistence enables multi-turn conversations, human-in-the-loop, time travel, and crash recovery.

Choosing a Backend

Agent Server note: When using LangGraph Agent Server, checkpointers are configured automatically — no manual setup needed.

Python Setup

# Development
from langgraph.checkpoint.memory import InMemorySaver
graph = builder.compile(checkpointer=InMemorySaver())

# Production (PostgreSQL)
from langgraph.checkpoint.postgres import PostgresSaver

DB_URI = "postgresql://user:pass@host:5432/db"
with PostgresSaver.from_conn_string(DB_URI) as checkpointer:
    # checkpointer.setup()  # Run once for initial schema
    graph = builder.compile(checkpointer=checkpointer)

    result = graph.invoke(
        {"messages": [{"role": "user", "content": "Hi"}]},
        {"configurable": {"thread_id": "session-1"}}
    )

TypeScript Setup

// Development
import { MemorySaver } from "@langchain/langgraph";
const graph = builder.compile({ checkpointer: new MemorySaver() });

// Production (PostgreSQL)
import { PostgresSaver } from "@langchain/langgraph-checkpoint-postgres";
const checkpointer = PostgresSaver.fromConnString(DB_URI);
// await checkpointer.setup();  // Run once
const graph = builder.compile({ checkpointer });

Thread Management

Every invocation requires a thread_id to identify the conversation:

config = {"configurable": {"thread_id": "user-123-session-1"}}
result = graph.invoke({"messages": [...]}, config)

Subgraph Persistence

Provide the checkpointer only on the parent graph — LangGraph propagates it to subgraphs automatically:

parent_graph = parent_builder.compile(checkpointer=checkpointer)
# Subgraphs inherit the checkpointer

To give a subgraph its own separate memory:

subgraph = sub_builder.compile(checkpointer=True)

For backend-specific configuration, migration between backends, and TTL settings, see references/persistence-backends.md.

State Typing

Python: TypedDict (Recommended)

from typing import TypedDict, Annotated, Literal

class AgentState(TypedDict):
    messages: Annotated[list[BaseMessage], add_messages]
    next: Literal["agent1", "agent2", "FINISH"]
    context: dict

Note: create_agent state schemas support TypedDict for custom agent state. Prefer TypedDict for agent state extensions.

TypeScript: StateSchema with Zod

import { StateSchema, MessagesValue, ReducedValue, UntrackedValue } from "@langchain/langgraph";
import { z } from "zod/v4";

const AgentState = new StateSchema({
  messages: MessagesValue,
  currentStep: z.string(),
  retryCount: z.number().default(0),

  // Custom reducer
  allSteps: new ReducedValue(
    z.array(z.string()).default(() => []),
    { inputSchema: z.string(), reducer: (current, newStep) => [...current, newStep] }
  ),

  // Transient state (not checkpointed)
  tempCache: new UntrackedValue(z.record(z.string(), z.unknown())),
});

// Extract types for use outside the graph builder
type State = typeof AgentState.State;
type Update = typeof AgentState.Update;

For Pydantic validation, advanced type patterns, and migration from untyped state, see references/state-typing.md.

Validation and Debugging

Validate State Schema

Run the validation script to check schema structure:

uv run scripts/validate_state_schema.py my_agent/state.py:MyState --verbose

Checks for: schema parsing issues, empty schemas, reducer annotation problems, message fields without reducers, routing fields without Literal types, and unsupported/unclear schema class patterns.

Test Reducers

Test reducer functions for correctness and edge cases:

uv run scripts/test_reducers.py my_agent/reducers.py:extend_list --verbose

Tests: basic merge, empty inputs, None handling, type consistency, nested structures, large inputs.

Inspect Checkpoints

Debug state evolution by inspecting saved checkpoints:

# List recent checkpoints
uv run scripts/inspect_checkpoints.py ./checkpoints.db

# Inspect specific checkpoint
uv run scripts/inspect_checkpoints.py ./checkpoints.db --checkpoint-id abc123 --thread-id thread-1

# View full history for a thread
uv run scripts/inspect_checkpoints.py ./checkpoints.db --thread-id thread-1 --history

inspect_checkpoints.py accepts either a direct SQLite DB path or a directory containing checkpoints.db.

Migrate Persisted State

When state shape changes require updating persisted checkpoint values:

# Dry run first
uv run scripts/migrate_state.py ./checkpoints.db migrations/add_field.py --dry-run

# Apply migration
uv run scripts/migrate_state.py ./checkpoints.db migrations/add_field.py

Migration script format:

def migrate(old_state: dict) -> dict:
    new_state = old_state.copy()
    new_state["new_field"] = "default_value"    # Add field
    new_state.pop("deprecated_field", None)      # Remove field
    return new_state

Common State Issues

For detailed debugging techniques, LangSmith tracing, and checkpoint inspection patterns, see references/state-debugging.md.

Resources

Scripts

References

State Templates