diff --git a/reme/agent/chat/fs_cli.py b/reme/agent/chat/fs_cli.py
index 953f4f59..19640fe9 100644
--- a/reme/agent/chat/fs_cli.py
+++ b/reme/agent/chat/fs_cli.py
@@ -6,6 +6,7 @@ from pathlib import Path
 from ...core.enumeration import Role, ChunkEnum
 from ...core.op import BaseReactStream
 from ...core.schema import Message, StreamChunk
+from ...tool.fs import BashTool, LsTool, ReadTool, WriteTool, EditTool
 
 
 class FsCli(BaseReactStream):
@@ -52,7 +53,17 @@ class FsCli(BaseReactStream):
 
         # Summarize current conversation and save to memory files
         current_date = datetime.now().strftime("%Y-%m-%d")
-        summarizer = FsSummarizer(tools=self.tools, working_dir=self.working_dir)
+        summarizer = FsSummarizer(
+            tools=[
+                BashTool(cwd=self.working_dir),
+                LsTool(cwd=self.working_dir),
+                ReadTool(cwd=self.working_dir),
+                WriteTool(cwd=self.working_dir),
+                EditTool(cwd=self.working_dir),
+            ],
+            working_dir=self.working_dir,
+            language=self.language,
+        )
 
         result = await summarizer.call(
             messages=self.messages,
@@ -113,7 +124,7 @@ class FsCli(BaseReactStream):
             left_messages = cut_result.get("left_messages", [])
 
         # Step 2: Generate summary via Compactor
-        compactor = FsCompactor()
+        compactor = FsCompactor(language=self.language)
         summary_content = await compactor.call(
             messages_to_summarize=messages_to_summarize,
             turn_prefix_messages=turn_prefix_messages,
diff --git a/reme/agent/fs/fs_summarizer.py b/reme/agent/fs/fs_summarizer.py
index d85e2c2a..d1406462 100644
--- a/reme/agent/fs/fs_summarizer.py
+++ b/reme/agent/fs/fs_summarizer.py
@@ -13,8 +13,9 @@ from ...core.utils import format_messages
 class FsSummarizer(BaseReact):
     """Retrieve personal memories through vector search and history reading."""
 
-    def __init__(self, memory_dir: str = "memory", version: str = "default", **kwargs):
+    def __init__(self, working_dir: str, memory_dir: str = "memory", version: str = "default", **kwargs):
         super().__init__(**kwargs)
+        self.working_dir: str = working_dir
         self.memory_dir: str = memory_dir
         self.version: str = version
 
@@ -29,6 +30,7 @@ class FsSummarizer(BaseReact):
                     content=self.prompt_format(
                         "user_message_default",
                         conversation=format_messages(messages, add_index=False),
+                        working_dir=self.working_dir,
                         date=date_str,
                         memory_dir=self.memory_dir,
                     ),
diff --git a/reme/agent/fs/fs_summarizer.yaml b/reme/agent/fs/fs_summarizer.yaml
index dde01f80..477a79e8 100644
--- a/reme/agent/fs/fs_summarizer.yaml
+++ b/reme/agent/fs/fs_summarizer.yaml
@@ -10,24 +10,78 @@ user_message: |
   If nothing to store, reply with [SILENT].
 
 user_message_default: |
+  Pre-compaction memory flush turn.
+  The session is near auto-compaction; capture durable memories to disk.
+
+  Current Date: {date}
+  Working Dir: {working_dir}
+
+  Store durable memories now (use {memory_dir}/YYYY-MM-DD.md).
+
+  Workflow:
+  1. Use read_tool to read {memory_dir}/YYYY-MM-DD.md (if file doesn't exist, read_tool tool will return an error)
+  2. Intelligently merge new information with existing content (skip if file doesn't exist):
+     - Avoid duplicating information that's already recorded
+     - Enrich existing entries with new details when relevant
+     - Maintain chronological order when applicable
+  3. Write the updated content:
+     - Use edit_tool to update specific sections when possible
+     - Use write_tool to overwrite the entire file if major restructuring is needed
+  4. Create {memory_dir}/ if it doesn't exist
+
+  Principles:
+  - Always preserve timestamps, dates, and time-related context
+  - Only add truly new or enriching information
+  - Keep entries concise but complete
+  - If nothing meaningful to store, reply with [SILENT]
+
+
+user_message_default_zh: |
+  预压缩内存刷新轮次。
+  当前会话即将进入自动压缩阶段；请将持久化记忆捕获并写入磁盘。
+
+  当前日期：{date}
+  工作目录：{working_dir}
+
+  立即存储持久化记忆（使用路径 {memory_dir}/YYYY-MM-DD.md）。
+
+  工作流程：
+  1. 使用 read_tool 读取 {memory_dir}/YYYY-MM-DD.md（如文件不存在，read_tool 会返回错误提示）
+  2. 智能合并新信息与现有内容（若文件不存在则跳过合并）：
+     - 避免重复已记录的信息
+     - 在相关时丰富现有条目的新细节
+     - 在适用时保持时间顺序
+  3. 写入更新后的内容：
+     - 尽可能使用 edit_tool 更新特定部分
+     - 如需大幅重构则使用 write_tool 覆盖整个文件
+  4. 如 {memory_dir}/ 不存在则创建
+
+  原则：
+  - 始终保留时间戳、日期和时间相关上下文
+  - 仅添加真正新的或有丰富价值的信息
+  - 保持条目简洁但完整
+  - 若无有意义的内容可存储，请回复 [SILENT]
+
+
+user_message_v1: |
   <conversation>
   {conversation}
   </conversation>
 
   The conversation is about to be compacted. Please extract persistent memories to disk.
   Current Date: {date}
+  Working Dir: {working_dir}
 
   Execution Flow:
   1. Determine if the conversation contains information worth storing
      - If no: Reply with reason + [SILENT]
      - If yes: Continue to step 2
 
-  2. Check file {memory_dir}/YYYY-MM-DD.md (use actual date)
-     - File doesn't exist: Write new memories directly
-     - File exists:
-       a) Read existing content
-       b) Compare and identify new/updated information
-       c) Prefer edit_tool for precise additions (preserves existing content); write_tool overwrites entire file
+  2. Use Read tool to read {memory_dir}/YYYY-MM-DD.md (use actual date)
+     - If file doesn't exist (Read returns error): Write new memories directly
+     - If file exists:
+       a) Compare and identify new/updated information from the read content
+       c) Prefer `edit_tool` for precise additions (preserves existing content); `write_tool` overwrites entire file
        d) If no new information: Reply with explanation + [SILENT]
 
   Update Principles:
@@ -41,25 +95,25 @@ user_message_default: |
 
   Please store persistent memories, keeping entries concise and well-structured.
 
-user_message_default_zh: |
+user_message_v1_zh: |
   <conversation>
   {conversation}
   </conversation>
 
   conversation即将压缩，请提取持久性记忆存储至磁盘。
   当前日期：{date}
+  工作目录: {working_dir}
 
   执行流程：
   1. 判断对话是否包含值得存储的信息
      - 若无：回复原因 + [SILENT]
      - 若有：继续步骤 2
 
-  2. 检查文件 {memory_dir}/YYYY-MM-DD.md（使用实际日期）
-     - 文件不存在：直接写入新记忆
+  2. 使用 Read 工具读取 {memory_dir}/YYYY-MM-DD.md（使用实际日期）
+     - 文件不存在（Read 返回错误）：直接使用 `write_tool` 写入新记忆
      - 文件已存在：
-       a) 读取现有内容
-       b) 对比识别新增/更新信息
-       c) 优先使用 edit_tool 精准添加新信息（保留已有内容），write_tool 会覆盖整个文件
+       a) 从读取的内容中对比识别新增/更新信息
+       c) 优先使用 `edit_tool` 精准添加新信息（保留已有内容）
        d) 若无新信息：回复说明 + [SILENT]
 
   更新原则：
diff --git a/reme/config/default.yaml b/reme/config/default.yaml
index 1d4dd742..3a7e7073 100644
--- a/reme/config/default.yaml
+++ b/reme/config/default.yaml
@@ -20,8 +20,7 @@ flows:
 llms:
   default:
     backend: openai
-#    model_name: qwen3-30b-a3b-instruct-2507
-    model_name: qwen3-next-80b-a3b-thinking
+    model_name: qwen3-30b-a3b-instruct-2507
 #    model_name: qwen3-30b-a3b-thinking-2507
     request_interval: 1
 #    temperature: 0.0001
diff --git a/reme/config/fs.yaml b/reme/config/fs.yaml
index 7c4be922..2bcf0516 100644
--- a/reme/config/fs.yaml
+++ b/reme/config/fs.yaml
@@ -4,7 +4,8 @@ llms:
   default:
     backend: openai
 #    model_name: qwen3-30b-a3b-instruct-2507
-    model_name: qwen3-30b-a3b-thinking-2507
+#    model_name: qwen3-30b-a3b-thinking-2507
+    model_name: qwen3-235b-a22b-thinking-2507
     request_interval: 1
 #    temperature: 0.0001