diff --git a/src/word_stats.py b/src/word_stats.py
index 6a4e612..0462bee 100644
--- a/src/word_stats.py
+++ b/src/word_stats.py
@@ -7,9 +7,7 @@ class WordStats:
     def __init__(self, lemma_features, db):
         self.lemma_features = lemma_features
         self.db = db
-
-        self.all_words = 0
-        self.memoized_msd_merges = {}
+        self.all_words = None
 
         self.db.init("""CREATE TABLE UniqWords (
             uw_id INTEGER PRIMARY KEY, 
@@ -19,6 +17,7 @@ class WordStats:
             frequency int
             )""")
         self.db.init("CREATE TABLE WordCount (lemma varchar(64), msd0 char, frequency int)")
+        self.db.init("CREATE TABLE NumWords (id INTEGER PRIMARY KEY, n INTEGER)")
 
         self.db.init("CREATE INDEX lemma_msd_text_on_uw ON UniqWords (lemma, msd, text)")
         self.db.init("CREATE INDEX lemma_on_uw ON UniqWords (lemma)")
@@ -34,10 +33,12 @@ class WordStats:
                 self.db.execute("""INSERT INTO UniqWords (lemma, msd, text, frequency) 
                     VALUES (:lemma, :msd, :text, 1)""", params)
 
-        self.db.commit()
-        self.all_words += len(words)
+        self.db.execute("INSERT INTO NumWords (n) VALUES (?)", (len(words),))
 
     def num_all_words(self):
+        if self.all_words is None:
+            cur = self.db.execute("SELECT sum(n) FROM NumWords")
+            self.all_words = int(cur.fetchone()[0])
         return self.all_words
 
     def generate_renders(self):