diff --git a/web/agent_api.py b/web/agent_api.py
index 4e4b766..7851b68 100644
--- a/web/agent_api.py
+++ b/web/agent_api.py
@@ -17,6 +17,7 @@ import io
 import json
 import os
 import queue
+import re
 import sys
 import threading
 import uuid
@@ -93,15 +94,74 @@ def _bigrams(text):
     return {t[i:i + 2] for i in range(len(t) - 1)}
 
 
+# ---------- 经验质量门槛 ----------
+# 自进化经验只应保存「独立操作任务」的成功套路。多轮对话里用户的短句
+# （质疑/纠正/催促，如「继续啊」「你确定我是卡1吗」「还没有完成啊」）
+# 不是新任务——把执行出错被纠正的轮次存成经验会教坏后续任务。
+# 用启发式过滤（零成本，可解释），配方层再校验工具名真实性。
+
+# 任务性动词：命中任一视为有明确操作诉求（疑问/纠错短句一般不含它们）
+_TASK_VERBS = ("打开", "搜索", "查看", "找到", "截图", "输入", "点击", "点开",
+               "发送", "安装", "卸载", "下载", "启动", "停止", "关闭", "退出",
+               "登录", "切换", "设置", "删除", "清理", "复制", "粘贴", "读取",
+               "剪贴板", "长按", "滑动", "播放", "发布", "检查", "看看",
+               "帮我", "给我", "请", "拍张", "查一下")
+# 对话续语开头：几乎只出现在承接上一轮（「继续啊」「还有吗」）
+_CONTINUE_PREFIXES = ("继续", "还有", "然后呢", "再来", "快点", "刚才",
+                      "接着", "上一步")
+# 强质疑/纠错信号（不含「为什么」——「查一下为什么」是正当任务）
+_DOUBT_MARKS = ("你确定", "是不是", "不是吧", "不是吗", "怎么都", "怎么还",
+                "还没有", "没看到", "我说的是", "你听我说", "不对吧", "你又",
+                "重新来", "错了", "你说得", "你回答")
+# 全部真实 MCP 工具（配方里出现不存在的 de_* 说明蒸馏模型在编造，弃存）
+_KNOWN_TOOLS = frozenset({
+    "de_list_devices", "de_screenshot", "de_tap", "de_swipe", "de_ui_tree",
+    "de_tap_element", "de_read_clipboard", "de_wake", "de_press_key",
+    "de_open_app", "de_stop_app", "de_foreground_app", "de_type_text",
+    "de_set_clipboard", "de_sleep", "de_ocr", "de_tap_text", "de_list_apps",
+    "de_list_tasks"})
+
+
+def _qualify_experience(prompt, recipe):
+    """经验入库前质量门槛，返回 True=值得保存。
+
+    1) prompt 太短 / 纯续语开头 / 质疑纠错 → 非独立任务，弃
+    2) 疑问短句（≤30 字、以 吗/？ 结尾）且无任务动词 → 追问/反问，弃
+    3) 配方含不存在的 de_* 工具（蒸馏模型自由发挥）→ 弃
+    """
+    t = "".join(c for c in (prompt or "") if not c.isspace())
+    if len(t) < 6:
+        _log.info("经验弃存：prompt 过短「%s」", t[:20])
+        return False
+    if any(t.startswith(p) for p in _CONTINUE_PREFIXES):
+        _log.info("经验弃存：对话续语开头「%s」", t[:20])
+        return False
+    if any(m in t for m in _DOUBT_MARKS):
+        _log.info("经验弃存：质疑/纠错语气「%s」", t[:20])
+        return False
+    if (t.endswith("吗") or t.endswith("？") or t.endswith("?")) \
+            and len(t) <= 30 and not any(v in t for v in _TASK_VERBS):
+        _log.info("经验弃存：无操作诉求的追问「%s」", t[:20])
+        return False
+    for name in re.findall(r"de_[a-z_]+", recipe or ""):
+        if name not in _KNOWN_TOOLS:
+            _log.info("经验弃存：配方含不存在的工具 %s", name)
+            return False
+    return True
+
+
 def _find_experiences(prompt, limit=2, threshold=0.10):
-    """按 bigram 重叠检索相似历史经验（prompt 与任务描述的字符相似度）。"""
+    """按 bigram 重叠检索相似历史经验（prompt 与任务描述的字符相似度）。
+
+    命中的经验 hits+1（回写），让被反复参考的有效经验浮到前面。
+    """
     try:
         if _flask_app is None:
             return ""
         with _flask_app.app_context():
             _ensure_exp_table()
             rows = db.session.execute(db.text(
-                "SELECT task_prompt, recipe, hits FROM agent_experience "
+                "SELECT id, task_prompt, recipe, hits FROM agent_experience "
                 "WHERE recipe != '' ORDER BY hits DESC, id DESC LIMIT 50")).fetchall()
     except Exception:
         return ""
@@ -111,13 +171,24 @@ def _find_experiences(prompt, limit=2, threshold=0.10):
     if not cur:
         return ""
     scored = []
-    for task_prompt, recipe, hits in rows:
+    for row_id, task_prompt, recipe, hits in rows:
         sim = len(cur & _bigrams(task_prompt)) / len(cur)
         if sim >= threshold:
-            scored.append((sim, hits or 0, recipe))
+            scored.append((sim, hits or 0, recipe, row_id))
     scored.sort(key=lambda x: (-x[0], -x[1]))
+    if scored:
+        # hits 回写（尽力而为，失败不影响检索）
+        try:
+            with _flask_app.app_context():
+                for _sim, _hits, _recipe, row_id in scored:
+                    db.session.execute(db.text(
+                        "UPDATE agent_experience SET hits=hits+1 WHERE id=:i"),
+                        {"i": row_id})
+                db.session.commit()
+        except Exception:
+            pass
     parts = []
-    for sim, _hits, recipe in scored[:limit]:
+    for sim, _hits, recipe, _row_id in scored[:limit]:
         parts.append(f"- {recipe[:600]}")
     return "\n".join(parts)
 
@@ -365,12 +436,13 @@ def _agent_thread(run_id, prompt, serial, cfg):
             try:
                 recipe = _distill_experience(cfg, prompt, " -> ".join(tool_seq))
                 if recipe and "配方" not in recipe[:50]:
-                    if _save_experience(prompt, recipe, " -> ".join(tool_seq)):
-                        _log.info("经验已写入记忆库，随事件流提示")
-                        q.put(("step", {"tool": "🧠 经验记忆",
-                                        "args": "本轮操作已提炼为经验并写入记忆库"
-                                                "（下次相似任务会自动参考）",
-                                        "image": None}))
+                    if _qualify_experience(prompt, recipe):
+                        if _save_experience(prompt, recipe, " -> ".join(tool_seq)):
+                            _log.info("经验已写入记忆库，随事件流提示")
+                            q.put(("step", {"tool": "🧠 经验记忆",
+                                            "args": "本轮操作已提炼为经验并写入记忆库"
+                                                    "（下次相似任务会自动参考）",
+                                            "image": None}))
             except Exception as e:
                 _log.warning(f"经验保存异常: {e}")
         q.put(("done", {"answer": answer}))