diff --git a/src/memory/common/llms/__init__.py b/src/memory/common/llms/__init__.py
index 9255f63..b2d37cc 100644
--- a/src/memory/common/llms/__init__.py
+++ b/src/memory/common/llms/__init__.py
@@ -24,13 +24,6 @@ from memory.common.llms.base import (
 )
 from memory.common.llms.anthropic_provider import AnthropicProvider
 from memory.common.llms.openai_provider import OpenAIProvider
-from memory.common.llms.usage_tracker import (
-    InMemoryUsageTracker,
-    RateLimitConfig,
-    TokenAllowance,
-    UsageBreakdown,
-    UsageTracker,
-)
 from memory.common import tokens
 
 __all__ = [
@@ -49,11 +42,6 @@ __all__ = [
     "StreamEvent",
     "LLMSettings",
     "create_provider",
-    "InMemoryUsageTracker",
-    "RateLimitConfig",
-    "TokenAllowance",
-    "UsageBreakdown",
-    "UsageTracker",
 ]
 
 logger = logging.getLogger(__name__)
@@ -93,28 +81,3 @@ def truncate(content: str, target_tokens: int) -> str:
     if len(content) > target_chars:
         return content[:target_chars].rsplit(" ", 1)[0] + "..."
     return content
-
-
-# bla = 1
-# from memory.common.llms import *
-# from memory.common.llms.tools.discord import make_discord_tools
-# from memory.common.db.connection import make_session
-# from memory.common.db.models import *
-
-# model = "anthropic/claude-sonnet-4-5"
-# provider = create_provider(model=model)
-# with make_session() as session:
-#     bot = session.query(DiscordBotUser).first()
-#     server = session.query(DiscordServer).first()
-#     channel = server.channels[0]
-#     tools = make_discord_tools(bot, None, channel, model)
-
-# def demo(msg: str):
-#     messages = [
-#         Message(
-#             role=MessageRole.USER,
-#             content=msg,
-#         )
-#     ]
-#     for m in provider.stream_with_tools(messages, tools):
-#         print(m)
diff --git a/src/memory/common/llms/base.py b/src/memory/common/llms/base.py
index f62eb4d..709d810 100644
--- a/src/memory/common/llms/base.py
+++ b/src/memory/common/llms/base.py
@@ -12,6 +12,7 @@ from PIL import Image
 
 from memory.common import settings
 from memory.common.llms.tools import ToolCall, ToolDefinition, ToolResult
+from memory.common.llms.usage import UsageTracker, RedisUsageTracker
 
 logger = logging.getLogger(__name__)
 
@@ -204,7 +205,9 @@ class LLMSettings:
 class BaseLLMProvider(ABC):
     """Base class for LLM providers."""
 
-    def __init__(self, api_key: str, model: str):
+    def __init__(
+        self, api_key: str, model: str, usage_tracker: UsageTracker | None = None
+    ):
         """
         Initialize the LLM provider.
 
@@ -215,6 +218,7 @@ class BaseLLMProvider(ABC):
         self.api_key = api_key
         self.model = model
         self._client: Any = None
+        self.usage_tracker: UsageTracker = usage_tracker or RedisUsageTracker()
 
     @abstractmethod
     def _initialize_client(self) -> Any:
diff --git a/src/memory/common/llms/usage/__init__.py b/src/memory/common/llms/usage/__init__.py
new file mode 100644
index 0000000..fde43e1
--- /dev/null
+++ b/src/memory/common/llms/usage/__init__.py
@@ -0,0 +1,17 @@
+from memory.common.llms.usage.redis_usage_tracker import RedisUsageTracker
+from memory.common.llms.usage.usage_tracker import (
+    InMemoryUsageTracker,
+    RateLimitConfig,
+    TokenAllowance,
+    UsageBreakdown,
+    UsageTracker,
+)
+
+__all__ = [
+    "InMemoryUsageTracker",
+    "RateLimitConfig",
+    "RedisUsageTracker",
+    "TokenAllowance",
+    "UsageBreakdown",
+    "UsageTracker",
+]
diff --git a/src/memory/common/llms/usage/redis_usage_tracker.py b/src/memory/common/llms/usage/redis_usage_tracker.py
new file mode 100644
index 0000000..468e9da
--- /dev/null
+++ b/src/memory/common/llms/usage/redis_usage_tracker.py
@@ -0,0 +1,88 @@
+"""Redis-backed usage tracker implementation."""
+
+import json
+from typing import Any, Iterable, Protocol
+
+import redis
+
+from memory.common import settings
+from memory.common.llms.usage.usage_tracker import (
+    RateLimitConfig,
+    UsageState,
+    UsageTracker,
+)
+
+
+class RedisClientProtocol(Protocol):
+    def get(self, key: str) -> Any:  # pragma: no cover - Protocol definition
+        ...
+
+    def set(
+        self, key: str, value: Any
+    ) -> Any:  # pragma: no cover - Protocol definition
+        ...
+
+    def scan_iter(
+        self, match: str
+    ) -> Iterable[Any]:  # pragma: no cover - Protocol definition
+        ...
+
+
+class RedisUsageTracker(UsageTracker):
+    """Tracks LLM usage for providers and models using Redis for persistence."""
+
+    def __init__(
+        self,
+        configs: dict[str, RateLimitConfig],
+        default_config: RateLimitConfig | None = None,
+        *,
+        redis_client: RedisClientProtocol | None = None,
+        key_prefix: str | None = None,
+    ) -> None:
+        super().__init__(configs=configs, default_config=default_config)
+        if redis_client is None:
+            redis_client = redis.Redis(
+                host=settings.REDIS_HOST,
+                port=int(settings.REDIS_PORT),
+                db=int(settings.REDIS_DB),
+                decode_responses=False,
+            )
+        self._redis = redis_client
+        prefix = key_prefix or settings.LLM_USAGE_REDIS_PREFIX
+        self._key_prefix = prefix.rstrip(":")
+
+    def get_state(self, model: str) -> UsageState:
+        redis_key = self._format_key(model)
+        payload = self._redis.get(redis_key)
+        if not payload:
+            return UsageState()
+        if isinstance(payload, bytes):
+            payload = payload.decode()
+        return UsageState.from_payload(json.loads(payload))
+
+    def iter_state_items(self) -> Iterable[tuple[str, UsageState]]:
+        pattern = f"{self._key_prefix}:*"
+        for redis_key in self._redis.scan_iter(match=pattern):
+            key = self._ensure_str(redis_key)
+            payload = self._redis.get(key)
+            if not payload:
+                continue
+            if isinstance(payload, bytes):
+                payload = payload.decode()
+            state = UsageState.from_payload(json.loads(payload))
+            yield key[len(self._key_prefix) + 1 :], state
+
+    def save_state(self, model: str, state: UsageState) -> None:
+        redis_key = self._format_key(model)
+        self._redis.set(
+            redis_key, json.dumps(state.to_payload(), separators=(",", ":"))
+        )
+
+    def _format_key(self, model: str) -> str:
+        return f"{self._key_prefix}:{model}"
+
+    @staticmethod
+    def _ensure_str(value: Any) -> str:
+        if isinstance(value, bytes):
+            return value.decode()
+        return str(value)
diff --git a/src/memory/common/llms/usage_tracker.py b/src/memory/common/llms/usage/usage_tracker.py
similarity index 87%
rename from src/memory/common/llms/usage_tracker.py
rename to src/memory/common/llms/usage/usage_tracker.py
index 90b1fb6..fbf9750 100644
--- a/src/memory/common/llms/usage_tracker.py
+++ b/src/memory/common/llms/usage/usage_tracker.py
@@ -5,6 +5,7 @@ from collections.abc import Iterable
 from dataclasses import dataclass, field
 from datetime import datetime, timedelta, timezone
 from threading import Lock
+from typing import Any
 
 
 @dataclass(frozen=True)
@@ -45,6 +46,40 @@ class UsageState:
     lifetime_input_tokens: int = 0
     lifetime_output_tokens: int = 0
 
+    def to_payload(self) -> dict[str, Any]:
+        return {
+            "events": [
+                {
+                    "timestamp": event.timestamp.isoformat(),
+                    "input_tokens": event.input_tokens,
+                    "output_tokens": event.output_tokens,
+                }
+                for event in self.events
+            ],
+            "window_input_tokens": self.window_input_tokens,
+            "window_output_tokens": self.window_output_tokens,
+            "lifetime_input_tokens": self.lifetime_input_tokens,
+            "lifetime_output_tokens": self.lifetime_output_tokens,
+        }
+
+    @classmethod
+    def from_payload(cls, payload: dict[str, Any]) -> "UsageState":
+        events = deque(
+            UsageEvent(
+                timestamp=datetime.fromisoformat(event["timestamp"]),
+                input_tokens=event["input_tokens"],
+                output_tokens=event["output_tokens"],
+            )
+            for event in payload.get("events", [])
+        )
+        return cls(
+            events=events,
+            window_input_tokens=payload.get("window_input_tokens", 0),
+            window_output_tokens=payload.get("window_output_tokens", 0),
+            lifetime_input_tokens=payload.get("lifetime_input_tokens", 0),
+            lifetime_output_tokens=payload.get("lifetime_output_tokens", 0),
+        )
+
 
 @dataclass
 class TokenAllowance:
@@ -76,9 +111,7 @@ class UsageBreakdown:
 
 def split_model_key(model: str) -> tuple[str, str]:
     if "/" not in model:
-        raise ValueError(
-            "model must be formatted as '<provider>/<model_name>'"
-        )
+        raise ValueError("model must be formatted as '<provider>/<model_name>'")
 
     provider, model_name = model.split("/", maxsplit=1)
     if not provider or not model_name:
@@ -205,9 +238,7 @@ class UsageTracker:
             if config.max_output_tokens is None:
                 output_remaining = None
             else:
-                output_remaining = (
-                    config.max_output_tokens - state.window_output_tokens
-                )
+                output_remaining = config.max_output_tokens - state.window_output_tokens
 
             return TokenAllowance(
                 input_tokens=clamp_non_negative(input_remaining),
@@ -313,4 +344,3 @@ def clamp_non_negative(value: int | None) -> int | None:
     if value is None:
         return None
     return 0 if value < 0 else value
-
diff --git a/src/memory/common/settings.py b/src/memory/common/settings.py
index 842a32c..9ea8c64 100644
--- a/src/memory/common/settings.py
+++ b/src/memory/common/settings.py
@@ -38,6 +38,7 @@ CELERY_BROKER_TYPE = os.getenv("CELERY_BROKER_TYPE", "redis").lower()
 REDIS_HOST = os.getenv("REDIS_HOST", "redis")
 REDIS_PORT = os.getenv("REDIS_PORT", "6379")
 REDIS_DB = os.getenv("REDIS_DB", "0")
+LLM_USAGE_REDIS_PREFIX = os.getenv("LLM_USAGE_REDIS_PREFIX", "llm_usage")
 CELERY_BROKER_USER = os.getenv(
     "CELERY_BROKER_USER", "kb" if CELERY_BROKER_TYPE == "amqp" else ""
 )
diff --git a/src/memory/workers/tasks/discord.py b/src/memory/workers/tasks/discord.py
index 8f6cfb4..102fbd7 100644
--- a/src/memory/workers/tasks/discord.py
+++ b/src/memory/workers/tasks/discord.py
@@ -58,6 +58,13 @@ def call_llm(
     msgs: list[str] = [],
     allowed_tools: list[str] = [],
 ) -> str | None:
+    provider = create_provider(model=model)
+    if provider.usage_tracker.is_rate_limited(model):
+        logger.error(
+            f"Rate limited for model {model}: {provider.usage_tracker.get_usage_breakdown(model=model)}"
+        )
+        return None
+
     tools = make_discord_tools(
         message.recipient_user.system_user,
         message.from_user,
@@ -73,7 +80,6 @@ def call_llm(
     system_prompt += comm_channel_prompt(
         session, message.recipient_user, message.channel
     )
-    provider = create_provider(model=model)
     messages = previous_messages(
         session,
         message.recipient_user and message.recipient_user.id,
diff --git a/tests/memory/common/llms/test_usage_tracker.py b/tests/memory/common/llms/test_usage_tracker.py
index e0df8ce..dbb9bc1 100644
--- a/tests/memory/common/llms/test_usage_tracker.py
+++ b/tests/memory/common/llms/test_usage_tracker.py
@@ -1,7 +1,24 @@
 from datetime import datetime, timedelta, timezone
+from typing import Iterable
 
 import pytest
 
+try:
+    import redis  # noqa: F401  # pragma: no cover - optional test dependency
+except ModuleNotFoundError:  # pragma: no cover - import guard for test envs
+    import sys
+    from types import SimpleNamespace
+
+    class _RedisStub(SimpleNamespace):
+        class Redis:  # type: ignore[no-redef]
+            def __init__(self, *args: object, **kwargs: object) -> None:
+                raise ModuleNotFoundError(
+                    "The 'redis' package is required to use RedisUsageTracker"
+                )
+
+    sys.modules.setdefault("redis", _RedisStub())
+
+from memory.common.llms.redis_usage_tracker import RedisUsageTracker
 from memory.common.llms.usage_tracker import (
     InMemoryUsageTracker,
     RateLimitConfig,
@@ -9,6 +26,24 @@ from memory.common.llms.usage_tracker import (
 )
 
 
+class FakeRedis:
+    def __init__(self) -> None:
+        self._store: dict[str, str] = {}
+
+    def get(self, key: str) -> str | None:
+        return self._store.get(key)
+
+    def set(self, key: str, value: str) -> None:
+        self._store[key] = value
+
+    def scan_iter(self, match: str) -> Iterable[str]:
+        from fnmatch import fnmatch
+
+        for key in list(self._store.keys()):
+            if fnmatch(key, match):
+                yield key
+
+
 @pytest.fixture
 def tracker() -> InMemoryUsageTracker:
     config = RateLimitConfig(
@@ -25,6 +60,23 @@ def tracker() -> InMemoryUsageTracker:
     )
 
 
+@pytest.fixture
+def redis_tracker() -> RedisUsageTracker:
+    config = RateLimitConfig(
+        window=timedelta(minutes=1),
+        max_input_tokens=1_000,
+        max_output_tokens=2_000,
+        max_total_tokens=2_500,
+    )
+    return RedisUsageTracker(
+        {
+            "anthropic/claude-3": config,
+            "anthropic/haiku": config,
+        },
+        redis_client=FakeRedis(),
+    )
+
+
 @pytest.mark.parametrize(
     "window, kwargs",
     [
@@ -139,6 +191,22 @@ def test_is_rate_limited_when_only_output_exceeds_limit() -> None:
     assert tracker.is_rate_limited("openai/gpt-4o")
 
 
+def test_redis_usage_tracker_persists_state(redis_tracker: RedisUsageTracker) -> None:
+    now = datetime(2024, 1, 1, tzinfo=timezone.utc)
+    redis_tracker.record_usage("anthropic/claude-3", 100, 200, timestamp=now)
+    redis_tracker.record_usage("anthropic/haiku", 50, 75, timestamp=now)
+
+    allowance = redis_tracker.get_available_tokens("anthropic/claude-3", timestamp=now)
+    assert allowance is not None
+    assert allowance.input_tokens == 900
+
+    breakdown = redis_tracker.get_usage_breakdown()
+    assert breakdown["anthropic"]["claude-3"].window_output_tokens == 200
+
+    items = dict(redis_tracker.iter_state_items())
+    assert set(items.keys()) == {"anthropic/claude-3", "anthropic/haiku"}
+
+
 def test_usage_tracker_base_not_instantiable() -> None:
     class DummyTracker(UsageTracker):
         pass