durable-state-patterns

Durable State Patterns

Build agents that remember state across failures, restarts, and sessions.

Why Durable State?

Without durability:

• Long-running agents lose progress on crash
• Users can't resume conversations after timeout
• No audit trail of agent decisions
• Expensive recomputation on every restart

With durability:

• Resume from any checkpoint
• Survive infrastructure failures
• Debug by replaying history
• Share state across instances

LangGraph Checkpointing

Basic Setup

from langgraph.graph import StateGraph
from langgraph.checkpoint.memory import MemorySaver
from langgraph.checkpoint.sqlite import SqliteSaver
from langgraph.checkpoint.postgres import PostgresSaver

# In-memory (development only)
memory_checkpointer = MemorySaver()

# SQLite (single instance)
sqlite_checkpointer = SqliteSaver.from_conn_string("checkpoints.db")

# PostgreSQL (production, multi-instance)
postgres_checkpointer = PostgresSaver.from_conn_string(
    "postgresql://user:pass@localhost/db"
)

# Compile with checkpointer
app = workflow.compile(checkpointer=postgres_checkpointer)

Thread-Based State

from uuid import uuid4

# Each conversation gets a unique thread_id
thread_id = str(uuid4())

# First invocation
config = {"configurable": {"thread_id": thread_id}}
result1 = app.invoke(
    {"messages": [{"role": "user", "content": "My name is Alice"}]},
    config
)

# Later invocation (same thread = same state)
result2 = app.invoke(
    {"messages": [{"role": "user", "content": "What's my name?"}]},
    config
)
# Agent remembers: "Your name is Alice"

# Different thread = fresh state
other_config = {"configurable": {"thread_id": str(uuid4())}}
result3 = app.invoke(
    {"messages": [{"role": "user", "content": "What's my name?"}]},
    other_config
)
# Agent doesn't know: "I don't have that information"

State Schema Design

Versioned State

from typing import TypedDict, Annotated
import operator

class AgentStateV1(TypedDict):
    """Version 1 of agent state."""
    messages: Annotated[list, operator.add]
    user_id: str

class AgentStateV2(TypedDict):
    """Version 2 with preferences."""
    messages: Annotated[list, operator.add]
    user_id: str
    preferences: dict  # New field
    state_version: int  # Track version

def migrate_v1_to_v2(old_state: AgentStateV1) -> AgentStateV2:
    """Migrate old state to new schema."""
    return {
        **old_state,
        "preferences": {},  # Default value
        "state_version": 2
    }

Separate Concerns

class ConversationState(TypedDict):
    """Short-term: current conversation."""
    messages: Annotated[list, operator.add]
    current_task: str

class UserProfileState(TypedDict):
    """Long-term: persists across conversations."""
    user_id: str
    name: str
    preferences: dict
    history_summary: str

class FullAgentState(TypedDict):
    """Combined state."""
    conversation: ConversationState
    profile: UserProfileState

Memory Tiers

┌─────────────────────────────────────────────────────────────┐
│                    Memory Architecture                       │
├─────────────────────────────────────────────────────────────┤
│                                                              │
│   ┌─────────────────────────────────────────────────────┐   │
│   │             Working Memory (In-Graph)                │   │
│   │   Current messages, tool results, intermediate       │   │
│   │   Lifetime: Single invocation                        │   │
│   └─────────────────────────────────────────────────────┘   │
│                              │                               │
│                              ▼                               │
│   ┌─────────────────────────────────────────────────────┐   │
│   │            Short-Term Memory (Thread)                │   │
│   │   Conversation history, session context              │   │
│   │   Lifetime: Single conversation/session              │   │
│   └─────────────────────────────────────────────────────┘   │
│                              │                               │
│                              ▼                               │
│   ┌─────────────────────────────────────────────────────┐   │
│   │            Long-Term Memory (User/Entity)            │   │
│   │   User preferences, facts, relationship history      │   │
│   │   Lifetime: Permanent                                │   │
│   └─────────────────────────────────────────────────────┘   │
│                                                              │
└─────────────────────────────────────────────────────────────┘

Implementation

import redis
from datetime import datetime

class ThreeTierMemory:
    """Three-tier memory system."""

    def __init__(self, redis_client: redis.Redis, checkpointer):
        self.redis = redis_client  # Short-term
        self.checkpointer = checkpointer  # Working memory via LangGraph
        self.db = PostgresDB()  # Long-term

    # Working memory: handled by LangGraph state

    # Short-term: Redis with TTL
    def get_session(self, session_id: str) -> dict:
        data = self.redis.get(f"session:{session_id}")
        return json.loads(data) if data else {}

    def save_session(self, session_id: str, data: dict, ttl: int = 3600):
        self.redis.setex(
            f"session:{session_id}",
            ttl,
            json.dumps(data)
        )

    # Long-term: Persistent database
    def get_user_profile(self, user_id: str) -> dict:
        return self.db.query(
            "SELECT * FROM user_profiles WHERE user_id = %s",
            (user_id,)
        )

    def update_user_profile(self, user_id: str, updates: dict):
        self.db.execute(
            "UPDATE user_profiles SET data = data || %s WHERE user_id = %s",
            (json.dumps(updates), user_id)
        )

Checkpoint Management

Manual Checkpoints

from langgraph.checkpoint import Checkpoint

# Get current checkpoint
checkpoint = app.get_state(config)
print(f"Checkpoint ID: {checkpoint.config['configurable']['checkpoint_id']}")
print(f"State: {checkpoint.values}")

# List all checkpoints for a thread
history = list(app.get_state_history(config))
for checkpoint in history:
    print(f"{checkpoint.config['configurable']['checkpoint_id']}: {checkpoint.values}")

# Rewind to specific checkpoint
old_checkpoint_id = "some-checkpoint-id"
rewound_config = {
    "configurable": {
        "thread_id": thread_id,
        "checkpoint_id": old_checkpoint_id
    }
}
result = app.invoke(new_input, rewound_config)

Checkpoint Cleanup

async def cleanup_old_checkpoints(
    checkpointer,
    max_age_days: int = 30
):
    """Clean up checkpoints older than max_age_days."""
    cutoff = datetime.now() - timedelta(days=max_age_days)

    # Implementation depends on checkpointer
    # For Postgres:
    await checkpointer.conn.execute(
        "DELETE FROM checkpoints WHERE created_at < $1",
        cutoff
    )

Recovery Patterns

Automatic Retry

from tenacity import retry, stop_after_attempt, wait_exponential

class ResilientAgent:
    def __init__(self, app, checkpointer):
        self.app = app
        self.checkpointer = checkpointer

    @retry(
        stop=stop_after_attempt(3),
        wait=wait_exponential(multiplier=1, min=4, max=10)
    )
    async def invoke_with_retry(self, input_state: dict, config: dict):
        """Invoke with automatic retry on failure."""
        try:
            return await self.app.ainvoke(input_state, config)
        except Exception as e:
            # State is already checkpointed, will resume on retry
            logger.error(f"Invocation failed, retrying: {e}")
            raise

Resume from Failure

async def resume_or_start(
    app,
    thread_id: str,
    input_state: dict
) -> dict:
    """Resume existing thread or start new one."""
    config = {"configurable": {"thread_id": thread_id}}

    # Check for existing state
    existing = app.get_state(config)

    if existing and existing.values:
        # Resume from checkpoint
        logger.info(f"Resuming thread {thread_id}")
        return await app.ainvoke(None, config)  # None = continue from checkpoint
    else:
        # Start fresh
        logger.info(f"Starting new thread {thread_id}")
        return await app.ainvoke(input_state, config)

Storage Backend Comparison

Backend	Use Case	Pros	Cons
MemorySaver	Development	Fast, simple	Lost on restart
SqliteSaver	Single instance	Persistent, simple	No concurrency
PostgresSaver	Production	Scalable, durable	Setup complexity
RedisSaver	High throughput	Fast, distributed	Memory limits

Best Practices

1. Choose thread_id wisely - User ID, session ID, or conversation ID
2. Version your state schema - Plan for migrations
3. Set checkpoint limits - Don't keep infinite history
4. Test recovery - Simulate failures in staging
5. Monitor checkpoint size - Large state = slow operations
6. Separate concerns - Working vs short-term vs long-term

Production Configuration

from langgraph.checkpoint.postgres import PostgresSaver
import asyncpg

async def create_production_checkpointer():
    """Create production-ready checkpointer."""

    # Connection pool for concurrency
    pool = await asyncpg.create_pool(
        "postgresql://user:pass@localhost/db",
        min_size=5,
        max_size=20
    )

    checkpointer = PostgresSaver(pool)

    # Initialize tables
    await checkpointer.setup()

    return checkpointer

# Use in app
checkpointer = await create_production_checkpointer()
app = workflow.compile(checkpointer=checkpointer)