diff --git a/langgraph/checkpoint/redis/aio.py b/langgraph/checkpoint/redis/aio.py
index 34538bf..65d72ff 100644
--- a/langgraph/checkpoint/redis/aio.py
+++ b/langgraph/checkpoint/redis/aio.py
@@ -33,6 +33,7 @@
     EMPTY_ID_SENTINEL,
     from_storage_safe_id,
     from_storage_safe_str,
+    safely_decode,
     to_storage_safe_id,
     to_storage_safe_str,
 )
@@ -212,12 +213,14 @@ async def aget_tuple(self, config: RunnableConfig) -> Optional[CheckpointTuple]:
             # Get the blob keys
             blob_key_pattern = f"{CHECKPOINT_BLOB_PREFIX}:{to_storage_safe_id(doc_thread_id)}:{to_storage_safe_str(doc_checkpoint_ns)}:*"
             blob_keys = await self._redis.keys(blob_key_pattern)
-            blob_keys = [key.decode() for key in blob_keys]
+            # Use safely_decode to handle both string and bytes responses
+            blob_keys = [safely_decode(key) for key in blob_keys]
 
             # Also get checkpoint write keys that should have the same TTL
             write_key_pattern = f"{CHECKPOINT_WRITE_PREFIX}:{to_storage_safe_id(doc_thread_id)}:{to_storage_safe_str(doc_checkpoint_ns)}:{to_storage_safe_id(doc_checkpoint_id)}:*"
             write_keys = await self._redis.keys(write_key_pattern)
-            write_keys = [key.decode() for key in write_keys]
+            # Use safely_decode to handle both string and bytes responses
+            write_keys = [safely_decode(key) for key in write_keys]
 
             # Apply TTL to checkpoint, blob keys, and write keys
             ttl_minutes = self.ttl_config.get("default_ttl")
@@ -895,9 +898,11 @@ async def _aload_pending_writes(
             None,
         )
         matching_keys = await self._redis.keys(pattern=writes_key)
+        # Use safely_decode to handle both string and bytes responses
+        decoded_keys = [safely_decode(key) for key in matching_keys]
         parsed_keys = [
-            BaseRedisSaver._parse_redis_checkpoint_writes_key(key.decode())
-            for key in matching_keys
+            BaseRedisSaver._parse_redis_checkpoint_writes_key(key)
+            for key in decoded_keys
         ]
         pending_writes = BaseRedisSaver._load_writes(
             self.serde,
diff --git a/langgraph/checkpoint/redis/ashallow.py b/langgraph/checkpoint/redis/ashallow.py
index da99885..58548c5 100644
--- a/langgraph/checkpoint/redis/ashallow.py
+++ b/langgraph/checkpoint/redis/ashallow.py
@@ -34,6 +34,7 @@
     REDIS_KEY_SEPARATOR,
     BaseRedisSaver,
 )
+from langgraph.checkpoint.redis.util import safely_decode
 
 SCHEMAS = [
     {
@@ -252,7 +253,9 @@ async def aput(
             # Process each existing blob key to determine if it should be kept or deleted
             if existing_blob_keys:
                 for blob_key in existing_blob_keys:
-                    key_parts = blob_key.decode().split(REDIS_KEY_SEPARATOR)
+                    # Use safely_decode to handle both string and bytes responses
+                    decoded_key = safely_decode(blob_key)
+                    key_parts = decoded_key.split(REDIS_KEY_SEPARATOR)
                     # The key format is checkpoint_blob:thread_id:checkpoint_ns:channel:version
                     if len(key_parts) >= 5:
                         channel = key_parts[3]
@@ -428,7 +431,8 @@ async def aget_tuple(self, config: RunnableConfig) -> Optional[CheckpointTuple]:
                 )
             )
             blob_keys = await self._redis.keys(blob_key_pattern)
-            blob_keys = [key.decode() for key in blob_keys]
+            # Use safely_decode to handle both string and bytes responses
+            blob_keys = [safely_decode(key) for key in blob_keys]
 
             # Apply TTL
             ttl_minutes = self.ttl_config.get("default_ttl")
@@ -554,7 +558,9 @@ async def aput_writes(
             # Process each existing writes key to determine if it should be kept or deleted
             if existing_writes_keys:
                 for write_key in existing_writes_keys:
-                    key_parts = write_key.decode().split(REDIS_KEY_SEPARATOR)
+                    # Use safely_decode to handle both string and bytes responses
+                    decoded_key = safely_decode(write_key)
+                    key_parts = decoded_key.split(REDIS_KEY_SEPARATOR)
                     # The key format is checkpoint_write:thread_id:checkpoint_ns:checkpoint_id:task_id:idx
                     if len(key_parts) >= 5:
                         key_checkpoint_id = key_parts[3]
@@ -700,9 +706,11 @@ async def _aload_pending_writes(
             thread_id, checkpoint_ns, checkpoint_id, "*", None
         )
         matching_keys = await self._redis.keys(pattern=writes_key)
+        # Use safely_decode to handle both string and bytes responses
+        decoded_keys = [safely_decode(key) for key in matching_keys]
         parsed_keys = [
-            BaseRedisSaver._parse_redis_checkpoint_writes_key(key.decode())
-            for key in matching_keys
+            BaseRedisSaver._parse_redis_checkpoint_writes_key(key)
+            for key in decoded_keys
         ]
         pending_writes = BaseRedisSaver._load_writes(
             self.serde,
diff --git a/langgraph/checkpoint/redis/base.py b/langgraph/checkpoint/redis/base.py
index f38f549..f227a33 100644
--- a/langgraph/checkpoint/redis/base.py
+++ b/langgraph/checkpoint/redis/base.py
@@ -18,6 +18,7 @@
 from langgraph.checkpoint.serde.types import ChannelProtocol
 
 from langgraph.checkpoint.redis.util import (
+    safely_decode,
     to_storage_safe_id,
     to_storage_safe_str,
 )
@@ -509,9 +510,12 @@ def _load_pending_writes(
         # Cast the result to List[bytes] to help type checker
         matching_keys: List[bytes] = self._redis.keys(pattern=writes_key)  # type: ignore[assignment]
 
+        # Use safely_decode to handle both string and bytes responses
+        decoded_keys = [safely_decode(key) for key in matching_keys]
+
         parsed_keys = [
-            BaseRedisSaver._parse_redis_checkpoint_writes_key(key.decode())
-            for key in matching_keys
+            BaseRedisSaver._parse_redis_checkpoint_writes_key(key)
+            for key in decoded_keys
         ]
         pending_writes = BaseRedisSaver._load_writes(
             self.serde,
@@ -541,6 +545,9 @@ def _load_writes(
 
     @staticmethod
     def _parse_redis_checkpoint_writes_key(redis_key: str) -> dict:
+        # Ensure redis_key is a string
+        redis_key = safely_decode(redis_key)
+
         parts = redis_key.split(REDIS_KEY_SEPARATOR)
         # Ensure we have at least 6 parts
         if len(parts) < 6:
diff --git a/langgraph/checkpoint/redis/shallow.py b/langgraph/checkpoint/redis/shallow.py
index 93fc4c6..5190729 100644
--- a/langgraph/checkpoint/redis/shallow.py
+++ b/langgraph/checkpoint/redis/shallow.py
@@ -26,6 +26,7 @@
     REDIS_KEY_SEPARATOR,
     BaseRedisSaver,
 )
+from langgraph.checkpoint.redis.util import safely_decode
 
 SCHEMAS = [
     {
@@ -179,7 +180,9 @@ def put(
         # Process each existing blob key to determine if it should be kept or deleted
         if existing_blob_keys:
             for blob_key in existing_blob_keys:
-                key_parts = blob_key.decode().split(REDIS_KEY_SEPARATOR)
+                # Use safely_decode to handle both string and bytes responses
+                decoded_key = safely_decode(blob_key)
+                key_parts = decoded_key.split(REDIS_KEY_SEPARATOR)
                 # The key format is checkpoint_blob:thread_id:checkpoint_ns:channel:version
                 if len(key_parts) >= 5:
                     channel = key_parts[3]
@@ -387,7 +390,10 @@ def get_tuple(self, config: RunnableConfig) -> Optional[CheckpointTuple]:
                     thread_id, checkpoint_ns
                 )
             )
-            blob_keys = [key.decode() for key in self._redis.keys(blob_key_pattern)]
+            # Use safely_decode to handle both string and bytes responses
+            blob_keys = [
+                safely_decode(key) for key in self._redis.keys(blob_key_pattern)
+            ]
 
             # Apply TTL
             self._apply_ttl_to_keys(checkpoint_key, blob_keys)
@@ -524,7 +530,9 @@ def put_writes(
         # Process each existing writes key to determine if it should be kept or deleted
         if existing_writes_keys:
             for write_key in existing_writes_keys:
-                key_parts = write_key.decode().split(REDIS_KEY_SEPARATOR)
+                # Use safely_decode to handle both string and bytes responses
+                decoded_key = safely_decode(write_key)
+                key_parts = decoded_key.split(REDIS_KEY_SEPARATOR)
                 # The key format is checkpoint_write:thread_id:checkpoint_ns:checkpoint_id:task_id:idx
                 if len(key_parts) >= 5:
                     key_checkpoint_id = key_parts[3]
diff --git a/langgraph/checkpoint/redis/util.py b/langgraph/checkpoint/redis/util.py
index 2489b4a..c4087a3 100644
--- a/langgraph/checkpoint/redis/util.py
+++ b/langgraph/checkpoint/redis/util.py
@@ -5,8 +5,14 @@
 that is lexicographically sortable. Typically, checkpoints that need
 sentinel values are from the first run of the graph, so this should
 generally be correct.
+
+This module also includes utility functions for safely handling Redis responses,
+including handling bytes vs string responses depending on how the Redis client is
+configured with decode_responses.
 """
 
+from typing import Any, Dict, List, Optional, Set, Tuple, Union
+
 EMPTY_STRING_SENTINEL = "__empty__"
 EMPTY_ID_SENTINEL = "00000000-0000-0000-0000-000000000000"
 
@@ -81,3 +87,46 @@ def from_storage_safe_id(value: str) -> str:
         return ""
     else:
         return value
+
+
+def safely_decode(obj: Any) -> Any:
+    """
+    Safely decode Redis responses, handling both string and bytes types.
+
+    This is especially useful when working with Redis clients configured with
+    different decode_responses settings. It recursively processes nested
+    data structures (dicts, lists, tuples, sets).
+
+    Based on RedisVL's convert_bytes function (redisvl.redis.utils.convert_bytes)
+    but implemented directly to avoid runtime import issues and ensure consistent
+    behavior with sets and other data structures. See PR #34 and referenced
+    implementation: https://github.com/redis/redis-vl-python/blob/9f22a9ad4c2166af6462b007833b456448714dd9/redisvl/redis/utils.py#L20
+
+    Args:
+        obj: The object to decode. Can be a string, bytes, or a nested structure
+            containing strings/bytes (dict, list, tuple, set).
+
+    Returns:
+        The decoded object with all bytes converted to strings.
+    """
+    if obj is None:
+        return None
+    elif isinstance(obj, bytes):
+        try:
+            return obj.decode("utf-8")
+        except UnicodeDecodeError:
+            # If decoding fails, return the original bytes
+            return obj
+    elif isinstance(obj, str):
+        return obj
+    elif isinstance(obj, dict):
+        return {safely_decode(k): safely_decode(v) for k, v in obj.items()}
+    elif isinstance(obj, list):
+        return [safely_decode(item) for item in obj]
+    elif isinstance(obj, tuple):
+        return tuple(safely_decode(item) for item in obj)
+    elif isinstance(obj, set):
+        return {safely_decode(item) for item in obj}
+    else:
+        # For other types (int, float, bool, etc.), return as is
+        return obj
diff --git a/tests/test_decode_responses.py b/tests/test_decode_responses.py
new file mode 100644
index 0000000..92e3c0b
--- /dev/null
+++ b/tests/test_decode_responses.py
@@ -0,0 +1,148 @@
+"""Tests for Redis key decoding functionality."""
+
+import os
+import time
+import uuid
+from typing import Any, Dict, Optional
+
+import pytest
+from redis import Redis
+
+from langgraph.checkpoint.redis.util import safely_decode
+
+
+def test_safely_decode_basic_types():
+    """Test safely_decode function with basic type inputs."""
+    # Test with bytes
+    assert safely_decode(b"test") == "test"
+
+    # Test with string
+    assert safely_decode("test") == "test"
+
+    # Test with None
+    assert safely_decode(None) is None
+
+    # Test with other types
+    assert safely_decode(123) == 123
+    assert safely_decode(1.23) == 1.23
+    assert safely_decode(True) is True
+
+
+def test_safely_decode_nested_structures():
+    """Test safely_decode function with nested data structures."""
+    # Test with dictionary
+    assert safely_decode({b"key": b"value"}) == {"key": "value"}
+    assert safely_decode({b"key1": b"value1", "key2": 123}) == {
+        "key1": "value1",
+        "key2": 123,
+    }
+
+    # Test with nested dictionary
+    nested_dict = {b"outer": {b"inner": b"value"}}
+    assert safely_decode(nested_dict) == {"outer": {"inner": "value"}}
+
+    # Test with list
+    assert safely_decode([b"item1", b"item2"]) == ["item1", "item2"]
+
+    # Test with tuple
+    assert safely_decode((b"item1", b"item2")) == ("item1", "item2")
+
+    # Test with set
+    decoded_set = safely_decode({b"item1", b"item2"})
+    assert isinstance(decoded_set, set)
+    assert "item1" in decoded_set
+    assert "item2" in decoded_set
+
+    # Test with complex nested structure
+    complex_struct = {
+        b"key1": [b"list_item1", {b"nested_key": b"nested_value"}],
+        b"key2": (b"tuple_item", 123),
+        b"key3": {b"set_item1", b"set_item2"},
+    }
+    decoded = safely_decode(complex_struct)
+    assert decoded["key1"][0] == "list_item1"
+    assert decoded["key1"][1]["nested_key"] == "nested_value"
+    assert decoded["key2"][0] == "tuple_item"
+    assert decoded["key2"][1] == 123
+    assert isinstance(decoded["key3"], set)
+    assert "set_item1" in decoded["key3"]
+    assert "set_item2" in decoded["key3"]
+
+
+@pytest.mark.parametrize("decode_responses", [True, False])
+def test_safely_decode_with_redis(decode_responses: bool, redis_url):
+    """Test safely_decode function with actual Redis responses using TestContainers."""
+    r = Redis.from_url(redis_url, decode_responses=decode_responses)
+
+    try:
+        # Clean up before test to ensure a clean state
+        r.delete("test:string")
+        r.delete("test:hash")
+        r.delete("test:list")
+        r.delete("test:set")
+
+        # Set up test data
+        r.set("test:string", "value")
+        r.hset("test:hash", mapping={"field1": "value1", "field2": "value2"})
+        r.rpush("test:list", "item1", "item2", "item3")
+        r.sadd("test:set", "member1", "member2")
+
+        # Test string value
+        string_val = r.get("test:string")
+        decoded_string = safely_decode(string_val)
+        assert decoded_string == "value"
+
+        # Test hash value
+        hash_val = r.hgetall("test:hash")
+        decoded_hash = safely_decode(hash_val)
+        assert decoded_hash == {"field1": "value1", "field2": "value2"}
+
+        # Test list value
+        list_val = r.lrange("test:list", 0, -1)
+        decoded_list = safely_decode(list_val)
+        assert decoded_list == ["item1", "item2", "item3"]
+
+        # Test set value
+        set_val = r.smembers("test:set")
+        decoded_set = safely_decode(set_val)
+        assert isinstance(decoded_set, set)
+        assert "member1" in decoded_set
+        assert "member2" in decoded_set
+
+        # Test key fetching
+        keys = r.keys("test:*")
+        decoded_keys = safely_decode(keys)
+        assert sorted(decoded_keys) == sorted(
+            ["test:string", "test:hash", "test:list", "test:set"]
+        )
+
+    finally:
+        # Clean up after test
+        r.delete("test:string")
+        r.delete("test:hash")
+        r.delete("test:list")
+        r.delete("test:set")
+        r.close()
+
+
+def test_safely_decode_unicode_error_handling():
+    """Test safely_decode function with invalid UTF-8 bytes."""
+    # Create bytes that will cause UnicodeDecodeError
+    invalid_utf8 = b"\xff\xfe\xfd"
+
+    # Should return the original bytes if it can't be decoded
+    result = safely_decode(invalid_utf8)
+    assert result == invalid_utf8
+
+    # Test with mixed valid and invalid in a complex structure
+    mixed = {
+        b"valid": b"This is valid UTF-8",
+        b"invalid": invalid_utf8,
+        b"nested": [b"valid", invalid_utf8],
+    }
+
+    result = safely_decode(mixed)
+    assert result["valid"] == "This is valid UTF-8"
+    assert result["invalid"] == invalid_utf8
+    assert result["nested"][0] == "valid"
+    assert result["nested"][1] == invalid_utf8