- Changed default model from qwen3-30b-a3b-instruct-2507 to qwen-flash
- Updated evaluation tools to use EVALUATION_PROMPT_FOR_QUESTION instead of QUESTION2
- Added new PROMPT_MEMZERO_JSON2 configuration with context priority rules
- Modified llm_request_for_json to use qwen-flash as default model
- Updated halumem evaluation to specify qwen3-max model explicitly for certain requests
- Add new analyze_dataset_stats.py script for comprehensive HaluMem dataset analysis
- Include statistics for user sessions, dialogues, content lengths and chunk distributions
- Replace rate limiting with concurrency control in BaseLLM using semaphore mechanism
- Update configuration to use max_concurrency instead of max_rps and rps_window
- Modify dialogue formatting to include only user messages in evaluation
- Add percentile calculations and detailed content size distribution metrics
- Implement session splitting logic based on character length thresholds
- Provide per-user statistics and summary tables for dataset analysis
- Refactor BaseLLM to use internal _chat_impl and _stream_chat_impl methods
- Remove rate limiting locks and timestamps from LLM initialization
- Add command-line interface for dataset statistics analysis tool
- Lowered max_rps from 20 to 2 to prevent rate limiting issues
- Reduced rps_window from 10 to 1 for stricter request throttling
- Updated default configuration values for better performance stability
- Increased max_rps from 9 to 20 in default config
- Removed tqdm progress bar for session processing
- Implemented asyncio semaphore for session concurrency control
- Added parallel processing of sessions within each user
- Updated logging to show session count per user
- Changed user processing from concurrent to sequential
- Added completion tracking with progress indicators
- Modified output formatting for better readability
- Integrated tqdm library for progress tracking
- Added progress bar display for user sessions processing
- Implemented session-by-session progress updates with custom description
- Maintained existing session processing logic while adding visual feedback
- Preserved all original functionality including logging and data handling
- Add ReMeRetrieverV2 and ReMeSummarizerV2 components
- Implement new memory tools including AddMemoryDrafts, RetrieveMemories, UpdateMemories
- Create PersonalSummarizerV2 with three-step workflow for memory management
- Add simplified evaluation script for ReMe on HaluMem benchmark
- Update base memory agent with enhanced logging capabilities
- Introduce new prompt formats and evaluation methods for question answering
- Modify LLM utilities and message formatting with markdown header stripping option
