at least 128 reused tokens and a tenth of what a plain request would prefill
 static bool scr_pays_eval(const common_scr_plan & plan, int32_t n_tokens) {
-    return !plan.empty() && plan.n_reused() >= std::max(512, (n_tokens - plan.restore) / 10);
+    return !plan.empty() && plan.n_reused() >= std::max(128, (n_tokens - plan.restore) / 10);
 }
 
 static arm_stats compare(const std::vector<float> & ref, const std::vector<float> & arm, const std::vector<llama_token> & teacher, int32_t n_vocab) {
diff --git a/tools/server/server-context.cpp b/tools/server/server-context.cpp
index 6016f357..01e7eaa3 100644
--- a/tools/server/server-context.cpp
+++ b/tools/server/server-context.cpp
@@ -1840,8 +1840,8 @@ private:
 
                 SLT_TRC(slot, " - checking sim = %.3f (%zu/%zu) > %.3f\n", f_sim_cur, lcp_len, task.tokens.size(), slot_prompt_similarity);
 
-                // select the current slot if the criteria match
-                if (f_sim_cur > f_sim_best && f_sim_cur > slot_prompt_similarity) {
+                // select the current slot if the criteria match (or if it shares at least 32 tokens, e.g. system prompt)
+                if (f_sim_cur > f_sim_best && (f_sim_cur > slot_prompt_similarity || lcp_len >= 32)) {
                     f_sim_best = f_sim_cur;
 
                     ret = &slot;
@@ -1938,9 +1938,13 @@ private:
 
                 ret->prompt_save(*prompt_cache);
 
-                // a slot chosen for SCR keeps its state: the splice reuses it
+                // a slot chosen for SCR keeps its state: the splice reuses it.
+                // also keep the slot state if it already shares tokens with the task, so we don't
+                // discard warm prefix cache in VRAM when the prompt cache has nothing better to load.
                 if (!scr_keep && !ret->prompt_load(*prompt_cache, task.tokens, skip)) {
-                    ret->prompt_clear();
+                    if (ret->prompt.tokens.get_common_prefix(task.tokens) == 0) {
+                        ret->prompt_clear();
+                    }
                 }
 
                 prompt_cache->update();
@@ -2697,9 +2701,9 @@ private:
         return scr_plan(slot.prompt, scr_n_state(slot), tokens);
     }
 
-    // the plan pays when it reuses at least 512 tokens and a tenth of what a plain request would prefill
+    // the plan pays when it reuses at least 128 tokens and a tenth of what a plain request would prefill
     static bool scr_pays(const common_scr_plan & plan, int32_t n_tokens) {
-        return !plan.empty() && plan.n_reused() >= std::max(512, (n_tokens - plan.restore) / 10);
+        return !plan.empty() && plan.n_reused() >= std::max(128, (n_tokens - plan.restore) / 10);
     }
 
     // loads the prompt-cache entry (RAM or disk) a splice can reuse more of than the slot's own state into the slot.
diff --git a/tools/server/server-task.cpp b/tools/server/server-task.cpp
index f9a4b636..fac61f1b 100644
--- a/tools/server/server-task.cpp
+++ b/tools/server/server-task.cpp
@@ -2326,8 +2326,8 @@ bool server_prompt_cache::load(server_prompt & prompt, const server_tokens & tok
 
         SRV_TRC("   - prompt with length %7zu, lcp = %7d, f_keep = %.3f, f_sim = %.3f\n", it->prompt.tokens.size(), lcp_cur, f_keep_cur, f_sim_cur);
 
-        // don't trash large prompts
-        if (f_keep_cur < 0.25f) {
+        // don't trash large prompts unless the matching prefix is at least 32 tokens
+        if (f_keep_cur < 0.25f && lcp_cur < 32) {
             continue;
         }
 
@@ -2352,7 +2352,8 @@ bool server_prompt_cache::load(server_prompt & prompt, const server_tokens & tok
             const float f_keep_cur = float(lcp_cur) / e.tokens.size();
             const float f_sim_cur  = float(lcp_cur) / tokens_new.size();
 
-            if (f_keep_cur < 0.25f) {
+            // don't trash large prompts unless the matching prefix is at least 32 tokens
+            if (f_keep_cur < 0.25f && lcp_cur < 32) {
                 continue;
             }