diff --git a/tools/server/server-task.cpp b/tools/server/server-task.cpp
index fbec0049..eea5165f 100644
--- a/tools/server/server-task.cpp
+++ b/tools/server/server-task.cpp
@@ -2503,44 +2503,37 @@ bool server_prompt_cache::load_entry(std::list<server_prompt_cache_state>::itera
     const bool learned = is_learned(it->prompt.meta);
     const bool consume = !keep_entry && !learned;
 
-    {
-        auto & data = it->data.main;
-
-        const size_t size = data.size();
-        const size_t n = llama_state_seq_set_data_ext(ctx_tgt, data.data(), size, id_slot, 0);
-        if (n != size) {
-            SRV_ERR("failed to restore state with size %zu\n", size);
+    // an empty state (a learned checkpoint without its state) restores nothing; llama reports 0 bytes read for it,
+    // which would otherwise pass as a complete restore
+    if (it->data.main.empty()) {
+        SRV_WRN("prompt cache: entry %s (%s, %d tokens) has no state to restore\n", hex_id(it->prompt.meta.id).c_str(),
+                cache_source_name(it->prompt.meta.source), it->prompt.n_tokens());
+        return false;
+    }
 
+    {
+        const auto & data = it->data.main;
+        const size_t n = llama_state_seq_set_data_ext(ctx_tgt, data.data(), data.size(), id_slot, 0);
+        if (n != data.size()) {
+            SRV_ERR("failed to restore state with size %zu\n", data.size());
             return false;
         }
-
-        if (consume) {
-            data.clear();
-            data.shrink_to_fit();
-        }
     }
 
     {
-        auto & data = it->data.drft;
-
+        const auto & data = it->data.drft;
         if (!data.empty()) {
             GGML_ASSERT(ctx_dft);
 
-            const size_t size = data.size();
-            const size_t n = llama_state_seq_set_data_ext(ctx_dft, data.data(), size, id_slot, 0);
-            if (n != size) {
-                SRV_WRN("failed to restore state with size %zu\n", size);
-
+            const size_t n = llama_state_seq_set_data_ext(ctx_dft, data.data(), data.size(), id_slot, 0);
+            if (n != data.size()) {
+                SRV_WRN("failed to restore state with size %zu\n", data.size());
                 return false;
             }
-
-            if (consume) {
-                data.clear();
-                data.shrink_to_fit();
-            }
         }
     }
 
+    // consumed only once both restores succeeded: a failed draft restore must not leave an entry without its state
     if (consume) {
         prompt = std::move(it->prompt);
 
@@ -2725,7 +2718,7 @@ bool server_prompt_cache::family_class(uint64_t family, cache_class & out) const
 
 void server_prompt_cache::apply_family(cache_meta & m) const {
     cache_class cls;
-    if (m.source != CACHE_SOURCE_EXPLICIT && family_class(m.family, cls)) {
+    if (m.source != CACHE_SOURCE_EXPLICIT && !is_learned(m) && family_class(m.family, cls)) {
         m.cls    = cls;
         m.source = CACHE_SOURCE_FAMILY;
     }
@@ -2947,7 +2940,8 @@ bool server_prompt_cache::set_entry(uint64_t id, cache_class cls, int64_t ttl) {
     bool found = false;
     auto apply = [&](cache_meta & m) {
         m.cls       = cls;
-        m.source    = CACHE_SOURCE_EXPLICIT;
+        // a learned prefix stays one: only its class and ttl change
+        m.source    = is_learned(m) ? CACHE_SOURCE_LEARNED : CACHE_SOURCE_EXPLICIT;
         m.ttl       = ttl;
         m.last_used = std::max(m.last_used, unix_now());
         set_expiry(m);