diff --git a/findmy/accessory.py b/findmy/accessory.py
index 4e79da4..c58eda9 100644
--- a/findmy/accessory.py
+++ b/findmy/accessory.py
@@ -6,8 +6,10 @@
 
 from __future__ import annotations
 
+import bisect
 import logging
 from abc import ABC, abstractmethod
+from dataclasses import dataclass
 from datetime import datetime, timedelta, timezone
 from typing import TYPE_CHECKING, Literal, TypedDict, overload
 
@@ -390,13 +392,24 @@ def __eq__(self, other: object) -> bool:
         )
 
 
+@dataclass(frozen=True)
+class _CacheTier:
+    """Configuration for a cache tier."""
+
+    interval: int  # Cache every n'th key
+    max_size: int | None  # Maximum number of keys to cache in this tier (None = unlimited)
+
+
 class _AccessoryKeyGenerator(KeyGenerator[KeyPair]):
     """KeyPair generator. Uses the same algorithm internally as FindMy accessories do."""
 
-    # cache enough keys for an entire week.
-    # every interval'th key is cached.
-    _CACHE_SIZE = 4 * 24 * 7  # 4 keys / hour
-    _CACHE_INTERVAL = 1  # cache every key
+    # Define cache tiers: (interval, max_size)
+    # Tier 1: Cache every 4th key (1 hour), keep up to 672 keys (2 weeks at 15min intervals)
+    # Tier 2: Cache every 672nd key (1 week), unlimited
+    _CACHE_TIERS = (
+        _CacheTier(interval=4, max_size=672),
+        _CacheTier(interval=672, max_size=None),
+    )
 
     def __init__(
         self,
@@ -422,7 +435,9 @@ def __init__(
         self._initial_sk = initial_sk
         self._key_type = key_type
 
-        self._sk_cache: dict[int, bytes] = {}
+        # Multi-tier cache: dict + sorted indices per tier
+        self._sk_caches: list[dict[int, bytes]] = [{} for _ in self._CACHE_TIERS]
+        self._cache_indices: list[list[int]] = [[] for _ in self._CACHE_TIERS]
 
         self._iter_ind = 0
 
@@ -441,36 +456,68 @@ def key_type(self) -> KeyPairType:
         """The type of key this generator produces."""
         return self._key_type
 
+    def _find_best_cached_sk(self, ind: int) -> tuple[int, bytes]:
+        """Find the largest cached index smaller than ind across all tiers."""
+        best_ind = 0
+        best_sk = self._initial_sk
+
+        for indices, cache in zip(self._cache_indices, self._sk_caches, strict=True):
+            if not indices:
+                continue
+
+            # Use bisect to find the largest index < ind in O(log n)
+            pos = bisect.bisect_left(indices, ind)
+            if pos == 0:  # No cached index less than ind
+                continue
+
+            cached_ind = indices[pos - 1]
+            if cached_ind > best_ind:
+                best_ind = cached_ind
+                best_sk = cache[cached_ind]
+
+        return best_ind, best_sk
+
+    def _update_caches(self, ind: int, sk: bytes) -> None:
+        """Update all applicable cache tiers with the computed key."""
+        for tier_idx, tier in enumerate(self._CACHE_TIERS):
+            if ind % tier.interval != 0:
+                continue
+
+            cache = self._sk_caches[tier_idx]
+            indices = self._cache_indices[tier_idx]
+
+            # Add to cache if not already present
+            if ind in cache:
+                continue
+            cache[ind] = sk
+            bisect.insort(indices, ind)
+
+            # Evict if cache exceeds size limit
+            if tier.max_size is not None and len(cache) > tier.max_size:
+                # If adding a historical key, evict smallest index
+                # If adding a future key, evict largest
+                evict_ind = indices.pop(0 if indices and ind > indices[0] else -1)
+
+                del cache[evict_ind]
+
     def _get_sk(self, ind: int) -> bytes:
         if ind < 0:
             msg = "The key index must be non-negative"
             raise ValueError(msg)
 
-        # retrieve from cache
-        cached_sk = self._sk_cache.get(ind)
-        if cached_sk is not None:
-            return cached_sk
+        # Check all caches for exact match
+        for cache in self._sk_caches:
+            cached_sk = cache.get(ind)
+            if cached_sk is not None:
+                return cached_sk
 
-        # not in cache: find largest cached index smaller than ind (if exists)
-        start_ind: int = 0
-        cur_sk: bytes = self._initial_sk
-        for cached_ind in self._sk_cache:
-            if cached_ind < ind and cached_ind > start_ind:
-                start_ind = cached_ind
-                cur_sk = self._sk_cache[cached_ind]
+        # Find best starting point across all tiers
+        start_ind, cur_sk = self._find_best_cached_sk(ind)
 
-        # compute and update cache
+        # Compute from best cached position to target
         for cur_ind in range(start_ind + 1, ind + 1):
             cur_sk = crypto.x963_kdf(cur_sk, b"update", 32)
-
-            # insert intermediate result into cache and evict oldest entry if necessary
-            if cur_ind % self._CACHE_INTERVAL == 0:
-                self._sk_cache[cur_ind] = cur_sk
-
-                if len(self._sk_cache) > self._CACHE_SIZE:
-                    # evict oldest entry
-                    oldest_ind = min(self._sk_cache.keys())
-                    del self._sk_cache[oldest_ind]
+            self._update_caches(cur_ind, cur_sk)
 
         return cur_sk