diff --git a/server/MetasequoiaImeEngine b/server/MetasequoiaImeEngine
index 452dd531..626b0507 160000
--- a/server/MetasequoiaImeEngine
+++ b/server/MetasequoiaImeEngine
@@ -1 +1 @@
-Subproject commit 452dd53125430195a06509eb0e9786f1c10657eb
+Subproject commit 626b05079fa0d3a35eafc4a9a52ecaac1a04654c
diff --git a/server/src/ai/ai_assistant.cpp b/server/src/ai/ai_assistant.cpp
index bb9794b5..28990189 100644
--- a/server/src/ai/ai_assistant.cpp
+++ b/server/src/ai/ai_assistant.cpp
@@ -9,6 +9,7 @@
 #include <condition_variable>
 #include <mutex>
 #include <thread>
+#include <unordered_map>
 
 namespace
 {
@@ -21,6 +22,14 @@ std::atomic<uint64_t> g_generation{0};
 AiAssistant::Request g_latest;
 std::chrono::steady_clock::time_point g_last_input;
 AiAssistant::ApplyCallback g_callback;
+std::unordered_map<std::string, std::string> g_candidate_cache;
+
+std::string BuildCacheKey(const AiAssistant::Request &request)
+{
+    nlohmann::json segments = request.pinyin_segments;
+    return request.config.provider + "\n" + request.config.endpoint + "\n" + request.config.model + "\n" +
+           segments.dump();
+}
 
 size_t WriteResponse(char *data, size_t size, size_t count, void *user)
 {
@@ -139,6 +148,19 @@ void WorkerLoop()
         if (!g_running) break;
         observed = g_generation.load();
         auto request = g_latest;
+        const auto cached = g_candidate_cache.find(BuildCacheKey(request));
+        if (!request.identity.empty() && cached != g_candidate_cache.end())
+        {
+            const std::string candidate = cached->second;
+            const uint64_t cached_generation = observed;
+            lock.unlock();
+            OutputDebugStringA(fmt::format("[ai-assistant] cache hit: generation={}, pinyin={}, candidate={}\n",
+                                          cached_generation,
+                                          nlohmann::json(request.pinyin_segments).dump(), candidate).c_str());
+            if (g_running && g_generation.load() == cached_generation && g_callback)
+                g_callback(candidate, request.identity, cached_generation);
+            continue;
+        }
         auto target = g_last_input + kIdleDelay;
         while (g_running && g_cv.wait_until(lock, target, [&] { return !g_running || g_generation.load() != observed; }))
         {
@@ -151,7 +173,16 @@ void WorkerLoop()
         if (request.identity.empty()) continue;
         const std::string candidate = Fetch(request, observed);
         if (!candidate.empty() && g_generation.load() == observed && g_callback)
+        {
+            {
+                std::lock_guard cache_lock(g_mutex);
+                g_candidate_cache[BuildCacheKey(request)] = candidate;
+            }
+            OutputDebugStringA(fmt::format("[ai-assistant] cache stored: generation={}, pinyin={}, candidate={}\n",
+                                          observed, nlohmann::json(request.pinyin_segments).dump(),
+                                          candidate).c_str());
             g_callback(candidate, request.identity, observed);
+        }
     }
 }
 } // namespace
@@ -172,6 +203,7 @@ void Stop()
     g_running = false;
     g_cv.notify_all();
     if (g_worker.joinable()) g_worker.join();
+    { std::lock_guard lock(g_mutex); g_candidate_cache.clear(); }
     curl_global_cleanup();
 }
 void OnInputChanged(Request request)
diff --git a/server/src/ipc/event_listener.cpp b/server/src/ipc/event_listener.cpp
index 496d9152..8b5653ae 100644
--- a/server/src/ipc/event_listener.cpp
+++ b/server/src/ipc/event_listener.cpp
@@ -1828,7 +1828,8 @@ void ApplyCloudCandidate(const std::string &candidate, const std::string &pinyin
     Global::candidate_ui.items.insert(Global::candidate_ui.items.begin() + insert_index,
                                       WordItem(pinyin, candidate, 1, CandidateSource::CloudSuggestion));
     // 还需要更新一下 dictionary 中的 cache
-    g_inputSession->cache_dynamic_candidate(cloud_query_state.cache_key, candidate);
+    g_inputSession->cache_dynamic_candidate(cloud_query_state.cache_key, candidate,
+                                            CandidateSource::CloudSuggestion);
     // 标记一下，云候选已经被加进来了
     Global::cloud_candidate.added = true;
     Global::cloud_candidate.word = candidate;
@@ -1873,6 +1874,7 @@ void ApplyAiCandidate(const std::string &candidate, const std::string &identity,
     const auto query = g_inputSession->get_cloud_query_state();
     const size_t insert_index = std::min<size_t>(2, items.size());
     items.insert(items.begin() + insert_index, WordItem(identity, candidate, 1, CandidateSource::AiSuggestion));
+    g_inputSession->cache_dynamic_candidate(query.cache_key, candidate, CandidateSource::AiSuggestion);
     OutputDebugStringA(fmt::format("[ai-assistant] candidate inserted: generation={}, index={}, candidate={}, "
                                   "identity={}\n", generation, insert_index + 1, candidate, identity).c_str());
     Global::ai_candidate = {true, candidate, query.committed_pinyin};
diff --git a/server/src/session/engine_input_session.cpp b/server/src/session/engine_input_session.cpp
index b70988ee..00dbe4cf 100644
--- a/server/src/session/engine_input_session.cpp
+++ b/server/src/session/engine_input_session.cpp
@@ -279,10 +279,11 @@ int EngineInputSession::remove_candidate(std::string pinyin, std::string word)
     return session_.delete_by_pinyin_and_word(std::move(pinyin), std::move(word));
 }
 
-int EngineInputSession::cache_dynamic_candidate(const std::string &pinyin, const std::string &word)
+int EngineInputSession::cache_dynamic_candidate(const std::string &pinyin, const std::string &word,
+                                                CandidateSource source)
 {
-    const int cache_result = session_.cache_dynamic_candidate(pinyin, word);
-    (void)session_.cache_dynamic_candidate_for_current_request(word);
+    const int cache_result = session_.cache_dynamic_candidate(pinyin, word, source);
+    (void)session_.cache_dynamic_candidate_for_current_request(word, source);
     return cache_result;
 }
 
diff --git a/server/src/session/engine_input_session.h b/server/src/session/engine_input_session.h
index f56ef24c..08ab937b 100644
--- a/server/src/session/engine_input_session.h
+++ b/server/src/session/engine_input_session.h
@@ -34,7 +34,8 @@ class EngineInputSession : public IInputSession
     int store_user_phrase(std::string pinyin, std::string word) override;
     int pin_candidate(std::string pinyin, std::string word) override;
     int remove_candidate(std::string pinyin, std::string word) override;
-    int cache_dynamic_candidate(const std::string &pinyin, const std::string &word) override;
+    int cache_dynamic_candidate(const std::string &pinyin, const std::string &word,
+                                CandidateSource source) override;
     SelectionTransition advance_composition_after_selection(const std::string &selected_pinyin,
                                                             const std::string &selected_word) override;
     CloudQueryState get_cloud_query_state() const override;
diff --git a/server/src/session/input_session.h b/server/src/session/input_session.h
index df67293a..b473a3b4 100644
--- a/server/src/session/input_session.h
+++ b/server/src/session/input_session.h
@@ -61,7 +61,8 @@ class IInputSession
     virtual int store_user_phrase(std::string pinyin, std::string word) = 0;
     virtual int pin_candidate(std::string pinyin, std::string word) = 0;
     virtual int remove_candidate(std::string pinyin, std::string word) = 0;
-    virtual int cache_dynamic_candidate(const std::string &pinyin, const std::string &word) = 0;
+    virtual int cache_dynamic_candidate(const std::string &pinyin, const std::string &word,
+                                        CandidateSource source) = 0;
     virtual SelectionTransition advance_composition_after_selection(const std::string &selected_pinyin,
                                                                    const std::string &selected_word) = 0;
     virtual CloudQueryState get_cloud_query_state() const = 0;
diff --git a/server/src/session/shuangpin_input_session.cpp b/server/src/session/shuangpin_input_session.cpp
index 9de7cbd7..db4f3b93 100644
--- a/server/src/session/shuangpin_input_session.cpp
+++ b/server/src/session/shuangpin_input_session.cpp
@@ -140,9 +140,10 @@ int ShuangpinInputSession::remove_candidate(std::string pinyin, std::string word
     return dictionary_->delete_by_pinyin_and_word(std::move(pinyin), std::move(word));
 }
 
-int ShuangpinInputSession::cache_dynamic_candidate(const std::string &pinyin, const std::string &word)
+int ShuangpinInputSession::cache_dynamic_candidate(const std::string &pinyin, const std::string &word,
+                                                   CandidateSource source)
 {
-    return dictionary_->insert_word_to_cached_buffer_series(pinyin, word);
+    return dictionary_->insert_word_to_cached_buffer_series(pinyin, word, source);
 }
 
 IInputSession::SelectionTransition
diff --git a/server/src/session/shuangpin_input_session.h b/server/src/session/shuangpin_input_session.h
index a23aaccc..e64dec5c 100644
--- a/server/src/session/shuangpin_input_session.h
+++ b/server/src/session/shuangpin_input_session.h
@@ -34,7 +34,8 @@ class ShuangpinInputSession : public IInputSession
     int store_user_phrase(std::string pinyin, std::string word) override;
     int pin_candidate(std::string pinyin, std::string word) override;
     int remove_candidate(std::string pinyin, std::string word) override;
-    int cache_dynamic_candidate(const std::string &pinyin, const std::string &word) override;
+    int cache_dynamic_candidate(const std::string &pinyin, const std::string &word,
+                                CandidateSource source) override;
     SelectionTransition advance_composition_after_selection(const std::string &selected_pinyin,
                                                            const std::string &selected_word) override;
     CloudQueryState get_cloud_query_state() const override;