diff --git a/benchmark/halumem/eval_reme.py b/benchmark/halumem/eval_reme.py
index 9b3ab7e7..a68ac3b2 100644
--- a/benchmark/halumem/eval_reme.py
+++ b/benchmark/halumem/eval_reme.py
@@ -105,7 +105,7 @@ class FileManager:
 
     def __init__(self, base_dir: str):
         self.base_dir = Path(base_dir)
-        self.tmp_dir = self.base_dir / "tmp"
+        self.tmp_dir = self.base_dir
         self.tmp_dir.mkdir(parents=True, exist_ok=True)
 
     def get_user_dir(self, user_name: str) -> Path:
@@ -829,8 +829,8 @@ if __name__ == "__main__":
     parser.add_argument(
         "--eval_model_name",
         type=str,
-        # default="qwen3-max",
-        default="qwen3-235b-a22b-instruct-2507",
+        default="qwen3-max",
+        # default="qwen3-235b-a22b-instruct-2507",
         help="Model name for evaluation (default: qwen3-max)"
     )
 
diff --git a/reme/agent/memory/default/personal_retriever.py b/reme/agent/memory/default/personal_retriever.py
index 895a7d26..5b2fa88c 100644
--- a/reme/agent/memory/default/personal_retriever.py
+++ b/reme/agent/memory/default/personal_retriever.py
@@ -79,7 +79,7 @@ class PersonalRetriever(BaseMemoryAgent):
                         ref_memory_id_key="",
                     )
                     for n in self.retrieved_nodes
-                ]
+                ],
             )
 
         result["retrieved_nodes"] = self.retrieved_nodes
diff --git a/reme/agent/memory/default/personal_summarizer.yaml b/reme/agent/memory/default/personal_summarizer.yaml
index 45708fbf..f30dabf6 100644
--- a/reme/agent/memory/default/personal_summarizer.yaml
+++ b/reme/agent/memory/default/personal_summarizer.yaml
@@ -6,19 +6,28 @@ system_prompt_s1: |
   {context}
 
   ## Task
-  ### Step 1
+  ### Step 1: Create Memory Draft
   Create a memory draft in `add_draft_and_retrieve_similar_memory` based on the latest conversation.
-  Use actual names from the conversation (e.g., "Bob likes apples") instead of generic references (e.g., "user likes apples"). Always record memories with real names.
-  The tool will retrieve similar historical memories via vector search to help you consolidate the memory store in Step 2.
+  - Use actual names from the conversation (e.g., "Bob likes apples") instead of generic references (e.g., "user likes apples")
+  - Always record memories with real names
+  - The tool will retrieve similar historical memories via vector search to help you consolidate in Step 2
 
-  ### Step 2
-  Update the vector store using `update_memory`.
-  Remove redundant memories via `memory_ids_to_delete` and add new ones via `memories_to_add` with their message_time and memory_content.
-  Requirements:
-  - Extract only what's explicitly stated in the conversation—no inferences, assumptions, or fabrications.
-  - Preserve all relevant historical and new memories, capturing how things change over time within the same topic.
-  - Keep the memory store well-organized: group related memories together and eliminate redundancy.
-  - **CAUTION on memory_ids_to_delete**: Only delete memories that are truly redundant or completely superseded by new memories.
+  ### Step 2: Update Memory Store
+  Update the vector store using `update_memory` to keep it well-organized and consolidated:
+
+  **What to Delete** (via `memory_ids_to_delete`):
+  - Duplicate memories with identical or highly similar content
+  - Memories that should be merged into a single consolidated entry
+
+  **What to Add** (via `memories_to_add` with message_time and memory_content):
+  - For each topic with changes: add ONE consolidated memory that merges related information
+  - New distinct memories that don't overlap with existing ones
+  - Updated memories that capture the latest state while preserving temporal evolution
+
+  ## Requirements
+  - Extract only what's explicitly stated—no inferences, assumptions, or fabrications
+  - Preserve temporal evolution: capture how things change over time within the same topic
+  - Maintain organization: group related memories by topic and eliminate all redundancy
 
 user_message_s1: |
   Complete the task by following Step 1 and Step 2 in order
@@ -31,17 +40,26 @@ system_prompt_s2: |
   {context}
 
   ## Task
-  ### Step 1
+  ### Step 1: Create Profile Draft
   Create a profile draft in `add_draft_and_read_all_profiles` based on the latest conversation.
-  The tool will return all existing profiles to help you maintain the profile store in Step 2.
+  - The tool will return all existing profiles to help you maintain the profile store in Step 2
 
-  ### Step 2
-  Update the profile store using `update_profile`.
-  Remove conflicting or redundant entries profiles via `profile_ids_to_delete` and add new ones via `profiles_to_add` with their message_time, profile_key, and profile_value.
-  Requirements:
-  - Extract only what's explicitly stated in the conversation—no inferences, assumptions, or fabrications.
-  - Keep the profile store well-organized: group related profiles together and eliminate duplicates.
-  - **CAUTION on profile_ids_to_delete**: Only delete profiles that are explicitly contradicted.
+  ### Step 2: Update Profile Store
+  Update the profile store using `update_profile` to keep it well-organized and consolidated:
+
+  **What to Delete** (via `profile_ids_to_delete`):
+  - Duplicate profiles with identical keys or values
+  - Conflicting profiles that contradict the new information
+  - Profiles that should be merged into a single consolidated entry
+
+  **What to Add** (via `profiles_to_add` with message_time, profile_key, and profile_value):
+  - For each profile key with changes: add ONE consolidated profile that merges related information
+  - New distinct profiles that don't overlap with existing ones
+  - Updated profiles that capture the latest state
+
+  ## Requirements
+  - Extract only what's explicitly stated—no inferences, assumptions, or fabrications
+  - Maintain organization: group related profiles by key and eliminate all redundancy
 
 user_message_s2: |
   Complete the task by following Step 1 and Step 2 in order