diff --git a/bucket.go b/bucket.go
index ecde1a3..5be20f2 100644
--- a/bucket.go
+++ b/bucket.go
@@ -103,3 +103,10 @@ func (b *bucket) nextTokensTime(executionTime ntime.Time, limit Limit, n int64)
 func (b *bucket) retryAfter(executionTime ntime.Time, limit Limit, n int64) time.Duration {
 	return max(0, b.nextTokensTime(executionTime, limit, n).Sub(executionTime))
 }
+
+// isFull checks if the bucket is full, i.e. has all the tokens it can have.
+//
+// ⚠️ caller is responsible for locking appropriately
+func (b *bucket) isFull(executionTime ntime.Time, limit Limit) bool {
+	return b.time.BeforeOrEqual(executionTime.Add(-limit.period))
+}
diff --git a/bucket_test.go b/bucket_test.go
index dd7f8d9..b398ec3 100644
--- a/bucket_test.go
+++ b/bucket_test.go
@@ -576,3 +576,26 @@ func TestBucket_RetryAfter(t *testing.T) {
 		require.Equal(t, expected, retryAfter, "retryAfter should be durationPerToken when requesting 2 tokens after exactly one durationPerToken has passed")
 	})
 }
+
+func TestBucket_IsFull(t *testing.T) {
+	t.Parallel()
+
+	now := ntime.Now()
+	limit := NewLimit(11, time.Second)
+	bucket := newBucket(now, limit)
+
+	// Newly created buckets are full
+	require.True(t, bucket.isFull(now, limit), "bucket should be full initially")
+
+	bucket.consumeTokens(now, limit, 5)
+	require.Equal(t, int64(6), bucket.remainingTokens(now, limit), "bucket should have 6 tokens remaining after consuming 5 tokens")
+	require.False(t, bucket.isFull(now, limit), "bucket should not be full after consuming 5 tokens")
+
+	now = now.Add(limit.durationPerToken * 4)
+	require.Equal(t, int64(10), bucket.remainingTokens(now, limit), "bucket should have 10 tokens remaining after consuming 5 tokens and waiting for 4")
+	require.False(t, bucket.isFull(now, limit), "bucket should not be full after consuming 5 tokens and waiting for 4")
+
+	now = now.Add(limit.durationPerToken)
+	require.Equal(t, int64(11), bucket.remainingTokens(now, limit), "bucket should have 11 tokens remaining")
+	require.True(t, bucket.isFull(now, limit), "bucket should be full after consuming 1 token and waiting for 4 durationPerTokens")
+}
diff --git a/limiter.go b/limiter.go
index 27eb522..e835a31 100644
--- a/limiter.go
+++ b/limiter.go
@@ -1,5 +1,7 @@
 package rate
 
+import "github.com/clipperhouse/ntime"
+
 // Limiter is a rate limiter that can be used to limit the rate of requests to a given key.
 type Limiter[TInput any, TKey comparable] struct {
 	keyFunc    KeyFunc[TInput, TKey]
@@ -41,3 +43,29 @@ func (r *Limiter[TInput, TKey]) getLimits(input TInput) []Limit {
 	}
 	return limits
 }
+
+// GC deletes buckets that are full, i.e, buckets for which enough
+// time has passed that they are no longer relevant. A full bucket
+// and a non-existent bucket have the same semantics.
+//
+// Without GC, buckets (memory) will grow unbounded.
+//
+// This can be a moderately expensive operation, depending
+// on the number of buckets. If you want a cheaper operation,
+// see [Clear].
+func (r *Limiter[TInput, TKey]) GC() (deleted int64) {
+	return r.buckets.gc(ntime.Now)
+}
+
+// Clear deletes all buckets. This is semantically
+// equivalent to refilling all buckets.
+//
+// You would use this method for garbage collection
+// purposes, as the limiter's memory will grow unbounded
+// otherwise.
+//
+// See also the [GC] method, which is more selective, and
+// only deletes buckets that are no longer meaningful.
+func (r *Limiter[TInput, TKey]) Clear() {
+	r.buckets.m.Clear()
+}
diff --git a/limiter_allow_test.go b/limiter_allow_test.go
index d2b4242..57cbeec 100644
--- a/limiter_allow_test.go
+++ b/limiter_allow_test.go
@@ -151,7 +151,7 @@ func TestLimiter_Allow(t *testing.T) {
 			}
 
 			expected := buckets * len(limiter.limits)
-			require.Equal(t, limiter.buckets.count(), expected, "buckets should have persisted after allow")
+			require.Equal(t, int64(expected), limiter.buckets.count(), "buckets should have persisted after allow")
 		})
 	})
 
diff --git a/limiter_peek_test.go b/limiter_peek_test.go
index 884b11f..eb49e57 100644
--- a/limiter_peek_test.go
+++ b/limiter_peek_test.go
@@ -29,7 +29,7 @@ func TestLimiter_Peek_NeverPersists(t *testing.T) {
 	}
 
 	// no buckets should have been stored
-	require.Equal(t, limiter.buckets.count(), 0, "buckets should not persist after peeking")
+	require.Equal(t, int64(0), limiter.buckets.count(), "buckets should not persist after peeking")
 }
 
 func TestLimiter_Peek_SingleBucket(t *testing.T) {
diff --git a/limiters.go b/limiters.go
index 74a7fd4..18f90bd 100644
--- a/limiters.go
+++ b/limiters.go
@@ -12,3 +12,34 @@ func Combine[TInput any, TKey comparable](limiters ...*Limiter[TInput, TKey]) *L
 		limiters: limiters,
 	}
 }
+
+// GC deletes buckets that are full, i.e, buckets for which enough
+// time has passed that they are no longer relevant. A full bucket
+// and a non-existent bucket have the same semantics.
+//
+// Without GC, buckets (memory) will grow unbounded.
+//
+// This can be a moderately expensive operation, depending
+// on the number of buckets. If you want a cheaper operation,
+// see [Clear].
+func (r *Limiters[TInput, TKey]) GC() (deleted int64) {
+	for _, limiter := range r.limiters {
+		deleted += limiter.GC()
+	}
+	return deleted
+}
+
+// Clear deletes all buckets. This is semantically
+// equivalent to refilling all buckets.
+//
+// You would use this method for garbage collection
+// purposes, as the limiter's memory will grow unbounded
+// otherwise.
+//
+// See also the [GC] method, which is more selective, and
+// only deletes buckets that are no longer meaningful.
+func (r *Limiters[TInput, TKey]) Clear() {
+	for _, limiter := range r.limiters {
+		limiter.Clear()
+	}
+}
diff --git a/syncmap.go b/syncmap.go
index 9f9c240..99100f3 100644
--- a/syncmap.go
+++ b/syncmap.go
@@ -8,7 +8,8 @@ import (
 
 // bucketMap is a specialized sync.Map for storing buckets
 type bucketMap[TKey comparable] struct {
-	m sync.Map
+	m  sync.Map
+	mu sync.RWMutex
 }
 
 // bucketSpec is a key for the bucket map, which includes the limit and the user key.
@@ -47,11 +48,34 @@ func (bm *bucketMap[TKey]) load(userKey TKey, limit Limit) (*bucket, bool) {
 	return nil, false
 }
 
-func (bm *bucketMap[TKey]) count() int {
-	count := 0
+func (bm *bucketMap[TKey]) count() int64 {
+	count := int64(0)
 	bm.m.Range(func(_, _ any) bool {
 		count++
 		return true
 	})
 	return count
 }
+
+// gc deletes buckets that are full, which has the
+// same semantics as the bucket not existing
+func (bm *bucketMap[TKey]) gc(timeFunc func() ntime.Time) (deleted int64) {
+	bm.mu.Lock()
+	defer bm.mu.Unlock()
+
+	bm.m.Range(func(key, value any) bool {
+		// I wonder if there is a possibility of a race here, not sure.
+		// Thinking about the timing between getting the bucket from the map
+		// and locking the bucket. Maybe it's not a problem.
+		spec := key.(bucketSpec[TKey])
+		b := value.(*bucket)
+		b.mu.Lock()
+		if b.isFull(timeFunc(), spec.limit) {
+			bm.m.Delete(key)
+			deleted++
+		}
+		b.mu.Unlock()
+		return true
+	})
+	return deleted
+}
diff --git a/syncmap_test.go b/syncmap_test.go
index 5441a44..a2ba955 100644
--- a/syncmap_test.go
+++ b/syncmap_test.go
@@ -3,6 +3,7 @@ package rate
 import (
 	"fmt"
 	"sync"
+	"sync/atomic"
 	"testing"
 	"time"
 
@@ -108,7 +109,7 @@ func TestBucketMap_Count(t *testing.T) {
 	executionTime := ntime.Now()
 
 	// Empty map should have count 0
-	require.Equal(t, 0, bm.count(), "empty map should have count 0")
+	require.Equal(t, int64(0), bm.count(), "empty map should have count 0")
 
 	// Add buckets with different keys and limits
 	for i := range 50 {
@@ -116,7 +117,7 @@ func TestBucketMap_Count(t *testing.T) {
 		bm.loadOrStore(key, executionTime, limit1)
 	}
 
-	require.Equal(t, 50, bm.count(), "should have 50 buckets after adding 50 different keys")
+	require.Equal(t, int64(50), bm.count(), "should have 50 buckets after adding 50 different keys")
 
 	// Add buckets with same keys but different limits
 	for i := range 50 {
@@ -124,7 +125,7 @@ func TestBucketMap_Count(t *testing.T) {
 		bm.loadOrStore(key, executionTime, limit2)
 	}
 
-	require.Equal(t, 100, bm.count(), "should have 100 buckets after adding same keys with different limits")
+	require.Equal(t, int64(100), bm.count(), "should have 100 buckets after adding same keys with different limits")
 
 	// Adding duplicate key-limit combinations should not increase count
 	for i := range 25 {
@@ -132,7 +133,7 @@ func TestBucketMap_Count(t *testing.T) {
 		bm.loadOrStore(key, executionTime, limit1)
 	}
 
-	require.Equal(t, 100, bm.count(), "should still have 100 buckets after duplicate additions")
+	require.Equal(t, int64(100), bm.count(), "should still have 100 buckets after duplicate additions")
 }
 
 func TestBucketMap_ConcurrentAccess(t *testing.T) {
@@ -190,5 +191,108 @@ func TestBucketMap_DifferentKeyTypes(t *testing.T) {
 	bucket3 := bm.loadOrStore(43, executionTime, limit)
 	require.False(t, bucket1 == bucket3, "different int keys should have different buckets")
 
-	require.Equal(t, 2, bm.count(), "should have 2 buckets for 2 different int keys")
+	require.Equal(t, int64(2), bm.count(), "should have 2 buckets for 2 different int keys")
+}
+
+func TestBucketMap_GC(t *testing.T) {
+	t.Parallel()
+
+	t.Run("gc deletes full buckets", func(t *testing.T) {
+		var bm bucketMap[string]
+		const count int64 = 1000
+
+		limit := NewLimit(10, time.Second)
+		executionTime := ntime.Now()
+
+		// Create a bunch of older buckets
+		for i := range count {
+			key := fmt.Sprintf("key%d", i)
+			b := bm.loadOrStore(key, executionTime, limit)
+			// 1/4 of the buckets will not be full
+			if i%4 == 0 {
+				b.consumeTokens(executionTime, limit, 1)
+			}
+		}
+
+		require.Equal(t, count, bm.count(), "should have 1000 buckets after creating 1000 older buckets")
+
+		// GC should delete 750 buckets full buckets, and not the 250 that are not full
+		{
+			deleted := bm.gc(func() ntime.Time {
+				return executionTime
+			})
+			require.Equal(t, int64(750), deleted, "should have 750 deleted buckets, since 1/4 of the buckets are full")
+			require.Equal(t, int64(250), bm.count(), "should have 250 buckets remaining after GC")
+		}
+
+		// Time passes
+		executionTime = executionTime.Add(time.Second)
+
+		// All remaining buckets are full now
+		{
+			deleted := bm.gc(func() ntime.Time {
+				return executionTime
+			})
+			require.Equal(t, int64(250), deleted, "should have deleted the remaining 250 buckets")
+			require.Equal(t, int64(0), bm.count(), "should have 0 buckets remaining after all deletions")
+		}
+	})
+
+	t.Run("gc concurrent with reads and writes", func(t *testing.T) {
+		var bm bucketMap[string]
+		const count int64 = 1000
+
+		limit := NewLimit(10, time.Second)
+		executionTime := ntime.Now()
+
+		// Create a bunch of buckets concurrently
+
+		// Trying to get the timing right for a good test,
+		// since a slow system like GitHub Actions seems
+		// to take a while to launch goroutines.
+		signal := make(chan struct{})
+		launched := int64(0)
+
+		var wg sync.WaitGroup
+		for i := range count {
+			wg.Add(1)
+			go func(i int64) {
+				defer wg.Done()
+
+				time.Sleep(time.Millisecond)
+				key := fmt.Sprintf("key%d", i)
+				b := bm.loadOrStore(key, executionTime, limit)
+				b.mu.Lock()
+				// 1/5 of the buckets will not be full
+				if i%5 == 0 {
+					b.consumeTokens(executionTime, limit, 1)
+				}
+				b.mu.Unlock()
+
+				// Signal when we reach 100 launched goroutines
+				if atomic.AddInt64(&launched, 1) == 100 {
+					close(signal)
+				}
+			}(i)
+		}
+
+		// Wait for 100 goroutines to launch before running GC,
+		// try to induce some concurrency.
+
+		<-signal
+		bm.gc(func() ntime.Time {
+			return executionTime
+		})
+		wg.Wait()
+
+		// Expect that some, but not all, deletions have happened,
+		// since there was concurrent creation of buckets.
+		require.Less(t, bm.count(), count, "some deletions should have happened")
+
+		// Now delete the remaining buckets, without concurrency
+		bm.gc(func() ntime.Time {
+			return executionTime
+		})
+		require.Equal(t, int64(200), bm.count(), "should have 200 buckets after deletion and GC")
+	})
 }