diff --git a/java/src/com/android/inputmethod/latin/BinaryDictionary.java b/java/src/com/android/inputmethod/latin/BinaryDictionary.java
index 95e1340..1b57918 100644
--- a/java/src/com/android/inputmethod/latin/BinaryDictionary.java
+++ b/java/src/com/android/inputmethod/latin/BinaryDictionary.java
@@ -174,8 +174,8 @@
     private static native int getFormatVersionNative(long dict);
     private static native int getProbabilityNative(long dict, int[] word);
     private static native int getMaxProbabilityOfExactMatchesNative(long dict, int[] word);
-    private static native int getNgramProbabilityNative(long dict, int[] word0,
-            boolean isBeginningOfSentence, int[] word1);
+    private static native int getNgramProbabilityNative(long dict, int[][] prevWordCodePointArrays,
+            boolean[] isBeginningOfSentenceArray, int[] word);
     private static native void getWordPropertyNative(long dict, int[] word,
             boolean isBeginningOfSentence, int[] outCodePoints, boolean[] outFlags,
             int[] outProbabilityInfo, ArrayList<int[]> outBigramTargets,
@@ -186,7 +186,7 @@
     private static native void getSuggestionsNative(long dict, long proximityInfo,
             long traverseSession, int[] xCoordinates, int[] yCoordinates, int[] times,
             int[] pointerIds, int[] inputCodePoints, int inputSize, int[] suggestOptions,
-            int[] prevWordCodePointArray, boolean isBeginningOfSentence,
+            int[][] prevWordCodePointArrays, boolean[] isBeginningOfSentenceArray,
             int[] outputSuggestionCount, int[] outputCodePoints, int[] outputScores,
             int[] outputIndices, int[] outputTypes, int[] outputAutoCommitFirstWordConfidence,
             float[] inOutLanguageWeight);
@@ -194,10 +194,11 @@
             int[] shortcutTarget, int shortcutProbability, boolean isBeginningOfSentence,
             boolean isNotAWord, boolean isBlacklisted, int timestamp);
     private static native boolean removeUnigramEntryNative(long dict, int[] word);
-    private static native boolean addNgramEntryNative(long dict, int[] word0,
-            boolean isBeginningOfSentence, int[] word1, int probability, int timestamp);
-    private static native boolean removeNgramEntryNative(long dict, int[] word0,
-            boolean isBeginningOfSentence, int[] word1);
+    private static native boolean addNgramEntryNative(long dict,
+            int[][] prevWordCodePointArrays, boolean[] isBeginningOfSentenceArray,
+            int[] word, int probability, int timestamp);
+    private static native boolean removeNgramEntryNative(long dict,
+            int[][] prevWordCodePointArrays, boolean[] isBeginningOfSentenceArray, int[] word);
     private static native int addMultipleDictionaryEntriesNative(long dict,
             LanguageModelParam[] languageModelParams, int startIndex);
     private static native String getPropertyNative(long dict, String query);
@@ -290,8 +291,8 @@
                 getTraverseSession(sessionId).getSession(), inputPointers.getXCoordinates(),
                 inputPointers.getYCoordinates(), inputPointers.getTimes(),
                 inputPointers.getPointerIds(), session.mInputCodePoints, inputSize,
-                session.mNativeSuggestOptions.getOptions(), session.mPrevWordCodePointArrays[0],
-                session.mIsBeginningOfSentenceArray[0], session.mOutputSuggestionCount,
+                session.mNativeSuggestOptions.getOptions(), session.mPrevWordCodePointArrays,
+                session.mIsBeginningOfSentenceArray, session.mOutputSuggestionCount,
                 session.mOutputCodePoints, session.mOutputScores, session.mSpaceIndices,
                 session.mOutputTypes, session.mOutputAutoCommitFirstWordConfidence,
                 session.mInputOutputLanguageWeight);
@@ -359,8 +360,8 @@
                 new boolean[Constants.MAX_PREV_WORD_COUNT_FOR_N_GRAM];
         prevWordsInfo.outputToArray(prevWordCodePointArrays, isBeginningOfSentenceArray);
         final int[] wordCodePoints = StringUtils.toCodePointArray(word);
-        return getNgramProbabilityNative(mNativeDict, prevWordCodePointArrays[0],
-                isBeginningOfSentenceArray[0], wordCodePoints);
+        return getNgramProbabilityNative(mNativeDict, prevWordCodePointArrays,
+                isBeginningOfSentenceArray, wordCodePoints);
     }
 
     public WordProperty getWordProperty(final String word, final boolean isBeginningOfSentence) {
@@ -456,8 +457,8 @@
                 new boolean[Constants.MAX_PREV_WORD_COUNT_FOR_N_GRAM];
         prevWordsInfo.outputToArray(prevWordCodePointArrays, isBeginningOfSentenceArray);
         final int[] wordCodePoints = StringUtils.toCodePointArray(word);
-        if (!addNgramEntryNative(mNativeDict, prevWordCodePointArrays[0],
-                isBeginningOfSentenceArray[0], wordCodePoints, probability, timestamp)) {
+        if (!addNgramEntryNative(mNativeDict, prevWordCodePointArrays,
+                isBeginningOfSentenceArray, wordCodePoints, probability, timestamp)) {
             return false;
         }
         mHasUpdated = true;
@@ -474,8 +475,8 @@
                 new boolean[Constants.MAX_PREV_WORD_COUNT_FOR_N_GRAM];
         prevWordsInfo.outputToArray(prevWordCodePointArrays, isBeginningOfSentenceArray);
         final int[] wordCodePoints = StringUtils.toCodePointArray(word);
-        if (!removeNgramEntryNative(mNativeDict, prevWordCodePointArrays[0],
-                isBeginningOfSentenceArray[0], wordCodePoints)) {
+        if (!removeNgramEntryNative(mNativeDict, prevWordCodePointArrays,
+                isBeginningOfSentenceArray, wordCodePoints)) {
             return false;
         }
         mHasUpdated = true;
diff --git a/native/jni/com_android_inputmethod_latin_BinaryDictionary.cpp b/native/jni/com_android_inputmethod_latin_BinaryDictionary.cpp
index 6b4fb79..3add84a 100644
--- a/native/jni/com_android_inputmethod_latin_BinaryDictionary.cpp
+++ b/native/jni/com_android_inputmethod_latin_BinaryDictionary.cpp
@@ -178,7 +178,7 @@
         jlong proximityInfo, jlong dicTraverseSession, jintArray xCoordinatesArray,
         jintArray yCoordinatesArray, jintArray timesArray, jintArray pointerIdsArray,
         jintArray inputCodePointsArray, jint inputSize, jintArray suggestOptions,
-        jintArray prevWordCodePointsForBigrams, jboolean isBeginningOfSentence,
+        jobjectArray prevWordCodePointArrays, jbooleanArray isBeginningOfSentenceArray,
         jintArray outSuggestionCount, jintArray outCodePointsArray, jintArray outScoresArray,
         jintArray outSpaceIndicesArray, jintArray outTypesArray,
         jintArray outAutoCommitFirstWordConfidenceArray, jfloatArray inOutLanguageWeight) {
@@ -201,20 +201,11 @@
     int pointerIds[inputSize];
     const jsize inputCodePointsLength = env->GetArrayLength(inputCodePointsArray);
     int inputCodePoints[inputCodePointsLength];
-    const jsize prevWordCodePointsLength =
-            prevWordCodePointsForBigrams ? env->GetArrayLength(prevWordCodePointsForBigrams) : 0;
-    int prevWordCodePointsInternal[prevWordCodePointsLength];
-    int *prevWordCodePoints = nullptr;
     env->GetIntArrayRegion(xCoordinatesArray, 0, inputSize, xCoordinates);
     env->GetIntArrayRegion(yCoordinatesArray, 0, inputSize, yCoordinates);
     env->GetIntArrayRegion(timesArray, 0, inputSize, times);
     env->GetIntArrayRegion(pointerIdsArray, 0, inputSize, pointerIds);
     env->GetIntArrayRegion(inputCodePointsArray, 0, inputCodePointsLength, inputCodePoints);
-    if (prevWordCodePointsForBigrams) {
-        env->GetIntArrayRegion(prevWordCodePointsForBigrams, 0, prevWordCodePointsLength,
-                prevWordCodePointsInternal);
-        prevWordCodePoints = prevWordCodePointsInternal;
-    }
 
     const jsize numberOfOptions = env->GetArrayLength(suggestOptions);
     int options[numberOfOptions];
@@ -248,8 +239,8 @@
     float languageWeight;
     env->GetFloatArrayRegion(inOutLanguageWeight, 0, 1 /* len */, &languageWeight);
     SuggestionResults suggestionResults(MAX_RESULTS);
-    const PrevWordsInfo prevWordsInfo(prevWordCodePoints, prevWordCodePointsLength,
-            isBeginningOfSentence);
+    const PrevWordsInfo prevWordsInfo = JniDataUtils::constructPrevWordsInfo(env,
+            prevWordCodePointArrays, isBeginningOfSentenceArray);
     if (givenSuggestOptions.isGesture() || inputSize > 0) {
         // TODO: Use SuggestionResults to return suggestions.
         dictionary->getSuggestions(pInfo, traverseSession, xCoordinates, yCoordinates,
@@ -391,41 +382,38 @@
 }
 
 static bool latinime_BinaryDictionary_addNgramEntry(JNIEnv *env, jclass clazz, jlong dict,
-        jintArray word0, jboolean isBeginningOfSentence, jintArray word1, jint probability,
-        jint timestamp) {
+        jobjectArray prevWordCodePointArrays, jbooleanArray isBeginningOfSentenceArray,
+        jintArray word, jint probability, jint timestamp) {
     Dictionary *dictionary = reinterpret_cast<Dictionary *>(dict);
     if (!dictionary) {
         return false;
     }
-    jsize word0Length = env->GetArrayLength(word0);
-    int word0CodePoints[word0Length];
-    env->GetIntArrayRegion(word0, 0, word0Length, word0CodePoints);
-    jsize word1Length = env->GetArrayLength(word1);
-    int word1CodePoints[word1Length];
-    env->GetIntArrayRegion(word1, 0, word1Length, word1CodePoints);
+    const PrevWordsInfo prevWordsInfo = JniDataUtils::constructPrevWordsInfo(env,
+            prevWordCodePointArrays, isBeginningOfSentenceArray);
+    jsize wordLength = env->GetArrayLength(word);
+    int wordCodePoints[wordLength];
+    env->GetIntArrayRegion(word, 0, wordLength, wordCodePoints);
     const std::vector<int> bigramTargetCodePoints(
-            word1CodePoints, word1CodePoints + word1Length);
+            wordCodePoints, wordCodePoints + wordLength);
     // Use 1 for count to indicate the bigram has inputted.
     const BigramProperty bigramProperty(&bigramTargetCodePoints, probability,
             timestamp, 0 /* level */, 1 /* count */);
-    const PrevWordsInfo prevWordsInfo(word0CodePoints, word0Length, isBeginningOfSentence);
     return dictionary->addNgramEntry(&prevWordsInfo, &bigramProperty);
 }
 
 static bool latinime_BinaryDictionary_removeNgramEntry(JNIEnv *env, jclass clazz, jlong dict,
-        jintArray word0, jboolean isBeginningOfSentence, jintArray word1) {
+        jobjectArray prevWordCodePointArrays, jbooleanArray isBeginningOfSentenceArray,
+        jintArray word) {
     Dictionary *dictionary = reinterpret_cast<Dictionary *>(dict);
     if (!dictionary) {
         return false;
     }
-    jsize word0Length = env->GetArrayLength(word0);
-    int word0CodePoints[word0Length];
-    env->GetIntArrayRegion(word0, 0, word0Length, word0CodePoints);
-    jsize word1Length = env->GetArrayLength(word1);
-    int word1CodePoints[word1Length];
-    env->GetIntArrayRegion(word1, 0, word1Length, word1CodePoints);
-    const PrevWordsInfo prevWordsInfo(word0CodePoints, word0Length, isBeginningOfSentence);
-    return dictionary->removeNgramEntry(&prevWordsInfo, word1CodePoints, word1Length);
+    const PrevWordsInfo prevWordsInfo = JniDataUtils::constructPrevWordsInfo(env,
+            prevWordCodePointArrays, isBeginningOfSentenceArray);
+    jsize wordLength = env->GetArrayLength(word);
+    int wordCodePoints[wordLength];
+    env->GetIntArrayRegion(word, 0, wordLength, wordCodePoints);
+    return dictionary->removeNgramEntry(&prevWordsInfo, wordCodePoints, wordLength);
 }
 
 // Returns how many language model params are processed.
@@ -672,7 +660,7 @@
     },
     {
         const_cast<char *>("getSuggestionsNative"),
-        const_cast<char *>("(JJJ[I[I[I[I[II[I[IZ[I[I[I[I[I[I[F)V"),
+        const_cast<char *>("(JJJ[I[I[I[I[II[I[[I[Z[I[I[I[I[I[I[F)V"),
         reinterpret_cast<void *>(latinime_BinaryDictionary_getSuggestions)
     },
     {
@@ -687,7 +675,7 @@
     },
     {
         const_cast<char *>("getNgramProbabilityNative"),
-        const_cast<char *>("(J[IZ[I)I"),
+        const_cast<char *>("(J[[I[Z[I)I"),
         reinterpret_cast<void *>(latinime_BinaryDictionary_getNgramProbability)
     },
     {
@@ -713,12 +701,12 @@
     },
     {
         const_cast<char *>("addNgramEntryNative"),
-        const_cast<char *>("(J[IZ[III)Z"),
+        const_cast<char *>("(J[[I[Z[III)Z"),
         reinterpret_cast<void *>(latinime_BinaryDictionary_addNgramEntry)
     },
     {
         const_cast<char *>("removeNgramEntryNative"),
-        const_cast<char *>("(J[IZ[I)Z"),
+        const_cast<char *>("(J[[I[Z[I)Z"),
         reinterpret_cast<void *>(latinime_BinaryDictionary_removeNgramEntry)
     },
     {
diff --git a/native/jni/src/suggest/core/session/prev_words_info.h b/native/jni/src/suggest/core/session/prev_words_info.h
index 640f6a2..e350c69 100644
--- a/native/jni/src/suggest/core/session/prev_words_info.h
+++ b/native/jni/src/suggest/core/session/prev_words_info.h
@@ -25,7 +25,6 @@
 namespace latinime {
 
 // TODO: Support n-gram.
-// This class does not take ownership of any code point buffers.
 class PrevWordsInfo {
  public:
     // No prev word information.
@@ -33,21 +32,52 @@
         clear();
     }
 
+    PrevWordsInfo(PrevWordsInfo &&prevWordsInfo) {
+        for (size_t i = 0; i < NELEMS(mPrevWordCodePoints); ++i) {
+            mPrevWordCodePointCount[i] = prevWordsInfo.mPrevWordCodePointCount[i];
+            memmove(mPrevWordCodePoints[i], prevWordsInfo.mPrevWordCodePoints[i],
+                    sizeof(mPrevWordCodePoints[i][0]) * mPrevWordCodePointCount[i]);
+            mIsBeginningOfSentence[i] = prevWordsInfo.mIsBeginningOfSentence[i];
+        }
+    }
+
+    // Construct from previous words.
+    PrevWordsInfo(const int prevWordCodePoints[][MAX_WORD_LENGTH],
+            const int *const prevWordCodePointCount, const bool *const isBeginningOfSentence,
+            const size_t prevWordCount) {
+        clear();
+        for (size_t i = 0; i < std::min(NELEMS(mPrevWordCodePoints), prevWordCount); ++i) {
+            if (prevWordCodePointCount[i] < 0 || prevWordCodePointCount[i] > MAX_WORD_LENGTH) {
+                continue;
+            }
+            memmove(mPrevWordCodePoints[i], prevWordCodePoints[i],
+                    sizeof(mPrevWordCodePoints[i][0]) * prevWordCodePointCount[i]);
+            mPrevWordCodePointCount[i] = prevWordCodePointCount[i];
+            mIsBeginningOfSentence[i] = isBeginningOfSentence[i];
+        }
+    }
+
+    // Construct from a previous word.
     PrevWordsInfo(const int *const prevWordCodePoints, const int prevWordCodePointCount,
             const bool isBeginningOfSentence) {
         clear();
-        mPrevWordCodePoints[0] = prevWordCodePoints;
+        if (prevWordCodePointCount > MAX_WORD_LENGTH || !prevWordCodePoints) {
+            return;
+        }
+        memmove(mPrevWordCodePoints[0], prevWordCodePoints,
+                sizeof(mPrevWordCodePoints[0][0]) * prevWordCodePointCount);
         mPrevWordCodePointCount[0] = prevWordCodePointCount;
         mIsBeginningOfSentence[0] = isBeginningOfSentence;
     }
 
     bool isValid() const {
-        for (size_t i = 0; i < NELEMS(mPrevWordCodePoints); ++i) {
-            if (mPrevWordCodePointCount[i] > MAX_WORD_LENGTH) {
-                return false;
-            }
+        if (mPrevWordCodePointCount[0] > 0) {
+            return true;
         }
-        return true;
+        if (mIsBeginningOfSentence[0]) {
+            return true;
+        }
+        return false;
     }
 
     void getPrevWordsTerminalPtNodePos(
@@ -168,13 +198,12 @@
 
     void clear() {
         for (size_t i = 0; i < NELEMS(mPrevWordCodePoints); ++i) {
-            mPrevWordCodePoints[i] = nullptr;
             mPrevWordCodePointCount[i] = 0;
             mIsBeginningOfSentence[i] = false;
         }
     }
 
-    const int *mPrevWordCodePoints[MAX_PREV_WORD_COUNT_FOR_N_GRAM];
+    int mPrevWordCodePoints[MAX_PREV_WORD_COUNT_FOR_N_GRAM][MAX_WORD_LENGTH];
     int mPrevWordCodePointCount[MAX_PREV_WORD_COUNT_FOR_N_GRAM];
     bool mIsBeginningOfSentence[MAX_PREV_WORD_COUNT_FOR_N_GRAM];
 };
diff --git a/native/jni/src/utils/jni_data_utils.h b/native/jni/src/utils/jni_data_utils.h
index 3514aee..cb82d3c 100644
--- a/native/jni/src/utils/jni_data_utils.h
+++ b/native/jni/src/utils/jni_data_utils.h
@@ -21,6 +21,7 @@
 
 #include "defines.h"
 #include "jni.h"
+#include "suggest/core/session/prev_words_info.h"
 #include "suggest/core/policy/dictionary_header_structure_policy.h"
 #include "suggest/policyimpl/dictionary/header/header_read_write_utils.h"
 #include "utils/char_utils.h"
@@ -95,6 +96,37 @@
         }
     }
 
+    static PrevWordsInfo constructPrevWordsInfo(JNIEnv *env, jobjectArray prevWordCodePointArrays,
+            jbooleanArray isBeginningOfSentenceArray) {
+        int prevWordCodePoints[MAX_PREV_WORD_COUNT_FOR_N_GRAM][MAX_WORD_LENGTH];
+        int prevWordCodePointCount[MAX_PREV_WORD_COUNT_FOR_N_GRAM];
+        bool isBeginningOfSentence[MAX_PREV_WORD_COUNT_FOR_N_GRAM];
+        jsize prevWordsCount = env->GetArrayLength(prevWordCodePointArrays);
+        for (size_t i = 0; i < NELEMS(prevWordCodePoints); ++i) {
+            prevWordCodePointCount[i] = 0;
+            isBeginningOfSentence[i] = false;
+            if (prevWordsCount <= static_cast<int>(i)) {
+                continue;
+            }
+            jintArray prevWord = (jintArray)env->GetObjectArrayElement(prevWordCodePointArrays, i);
+            if (!prevWord) {
+                continue;
+            }
+            jsize prevWordLength = env->GetArrayLength(prevWord);
+            if (prevWordLength > MAX_WORD_LENGTH) {
+                continue;
+            }
+            env->GetIntArrayRegion(prevWord, 0, prevWordLength, prevWordCodePoints[i]);
+            prevWordCodePointCount[i] = prevWordLength;
+            jboolean isBeginningOfSentenceBoolean = JNI_FALSE;
+            env->GetBooleanArrayRegion(isBeginningOfSentenceArray, i, 1 /* len */,
+                    &isBeginningOfSentenceBoolean);
+            isBeginningOfSentence[i] = isBeginningOfSentenceBoolean == JNI_TRUE;
+        }
+        return PrevWordsInfo(prevWordCodePoints, prevWordCodePointCount, isBeginningOfSentence,
+                MAX_PREV_WORD_COUNT_FOR_N_GRAM);
+    }
+
     static void putBooleanToArray(JNIEnv *env, jbooleanArray array, const int index,
             const jboolean value) {
         env->SetBooleanArrayRegion(array, index, 1 /* len */, &value);