feat(memory): add workspace_id fallback support for memory config resolution

- Add workspace_id fallback parameter to memory config loading across all services
- Update hot_memory_tags.py to pass workspace_id when resolving memory configuration
- Enhance emotion_analytics_service.py to support workspace_id as fallback for config resolution
- Improve implicit_memory_service.py with workspace_id fallback in config loading
- Update memory_agent_service.py to handle workspace_id resolution and add refactoring TODO
- Enhance preference_analysis.jinja2 prompt with critical guidance on supporting_evidence extraction
- Add validation to check both config_id and workspace_id before raising configuration errors
- Improve error handling and logging for memory configuration resolution across services
- This enables more flexible memory configuration resolution when config_id is unavailable

api/app/core/memory/analytics/hot_memory_tags.py

@@ -39,16 +39,20 @@ async def filter_tags_with_llm(tags: List[str], end_user_id: str) -> List[str]:
             
             connected_config = get_end_user_connected_config(end_user_id, db)
             config_id = connected_config.get("memory_config_id")
+            workspace_id = connected_config.get("workspace_id")
             
-            if not config_id:
+            if not config_id and not workspace_id:
                 raise ValueError(
                     f"No memory_config_id found for end_user_id: {end_user_id}. "
                     "Please ensure the user has a valid memory configuration."
                 )
             
-            # Use the config_id to get the proper LLM client
+            # Use the config_id to get the proper LLM client with workspace fallback
             config_service = MemoryConfigService(db)
-            memory_config = config_service.load_memory_config(config_id)
+            memory_config = config_service.load_memory_config(
+                config_id=config_id,
+                workspace_id=workspace_id
+            )
             
             if not memory_config.llm_model_id:
                 raise ValueError(

api/app/core/memory/analytics/implicit_memory/prompts/preference_analysis.jinja2

@@ -16,6 +16,7 @@ Summary {{ loop.index }}:
 3. DO NOT use long phrases - use short nouns or noun phrases
 4. Only include preferences with confidence_score >= 0.3
 5. **IMPORTANT: Output language MUST match the input language. If summaries are in Chinese, output in Chinese. If in English, output in English.**
+6. **CRITICAL: supporting_evidence must be DIRECT QUOTES or paraphrases from the user's actual statements. DO NOT reference summary numbers (e.g., "Summary 1", "摘要1"). DO NOT describe what the summary contains. Extract the actual user behavior or statement as evidence.**
 
 ## Output Format
 {
@@ -38,6 +39,16 @@ Summary {{ loop.index }}:
   ]
 }
 
+## BAD supporting_evidence examples (DO NOT do this):
+- "Summary 1：西湖为核心景区" ❌
+- "摘要2中提到喜欢咖啡" ❌
+- "Based on Summary 3" ❌
+
+## GOOD supporting_evidence examples:
+- "去过西湖断桥、苏堤" ✓
+- "每天早上喝咖啡" ✓
+- "mentioned visiting the lake twice" ✓
+
 ## Example (English input → English output)
 {
   "preferences": [