diff --git a/docs/devir-notlari.md b/docs/devir-notlari.md
index e7b196b..1585782 100644
--- a/docs/devir-notlari.md
+++ b/docs/devir-notlari.md
@@ -20,6 +20,7 @@ All jobs are detached (parent PID 1). They do not stop when the Claude Code sess
 | DeepSeek filter W6A (last task G0125) | `xargs python3 -m harness.empirical ... --run-base 15000` | 57246 / 57253 | `runs/emp_deepseek6a.log` | about 18:45 |
 | DeepSeek filter W7A (11 tasks, budget-60 reruns) | `xargs ... --run-base 17000 (list runs/stage1/w7a.txt)` | 59503 / 59510 | `runs/emp_deepseek7a.log` | about 19:45 |
 | Night chain | `train/night_chain.sh` | 60017 | `runs/stage1/night_chain.log` | see below |
+| Server restart with prompt cache limit (after the T01 test; new server PID in the log) | `train/restart_server.sh` | 60447 | `runs/stage1/restart_server.log` | after the T01 test |
 
 Night chain (`train/night_chain.sh`):
 1. When W6A ends: starts W7B (13 DeepSeek reruns, list `runs/stage1/w7b.txt`, log `runs/emp_deepseek7b.log`,
diff --git a/train/restart_server.sh b/train/restart_server.sh
new file mode 100755
index 0000000..da74220
--- /dev/null
+++ b/train/restart_server.sh
@@ -0,0 +1,11 @@
+#!/bin/sh
+# Restart the MLX server with the prompt cache limit when the T01 test (PID 59414) has ended,
+# before the night chain starts the baseline. Log: runs/stage1/restart_server.log
+cd "$(dirname "$0")/.."
+while kill -0 59414 2>/dev/null; do sleep 20; done
+echo "$(date '+%F %T') T01 test ended; restart server"
+kill 59352; while kill -0 59352 2>/dev/null; do sleep 2; done
+nohup train/serve.sh > runs/stage1/server.log 2>&1 &
+echo "$(date '+%F %T') new server PID $!"
+i=0; until curl -s -m 2 http://127.0.0.1:8080/v1/models >/dev/null 2>&1 || [ $i -ge 60 ]; do sleep 5; i=$((i+1)); done
+echo "$(date '+%F %T') server ready: $(curl -s -m 2 http://127.0.0.1:8080/v1/models | head -c 80)"
diff --git a/train/serve.sh b/train/serve.sh
index f927369..148d701 100755
--- a/train/serve.sh
+++ b/train/serve.sh
@@ -1,11 +1,13 @@
 #!/bin/sh
 # Serve the base model (no adapter) with mlx_lm.server. Settings: train/README.md.
 # Usage: train/serve.sh [--adapter-path PATH]
+# Prompt cache limit: without it the cache grew to 7.9 GB in 15 min (2026-10-03); A4H needs 22 GB.
 cd "$(dirname "$0")/.."
 exec train/.venv/bin/mlx_lm.server \
   --model "$HOME/models/Qwen3.8-27B-4bit" \
   --host 127.0.0.1 --port 8080 \
   --temp 0.2 --top-p 0.95 --top-k 20 --min-p 0 \
   --max-tokens 32768 \
+  --prompt-cache-size 4 --prompt-cache-bytes 6000000000 \
   --chat-template-args '{"enable_thinking": true, "reasoning_effort": "medium"}' \
   "$@"