Skip to content

Latest commit

 

History

History
172 lines (147 loc) · 7.51 KB

CHANGELOG.md

File metadata and controls

172 lines (147 loc) · 7.51 KB

Changelog

All notable changes to the MateBot project.

[Unreleased]

[1.1.0] - 2025-02-01

Changed

  • Fully automatic memory mode - Every message now automatically includes auto-memory instructions
  • Auto-memory instruction injected into every prompt via _get_or_init_auto_memory_instruction()
  • CCL-style memory format - Migrated from XML tags to elegant key-value format
    • Format: -- memory\nkey = value\n-- (inspired by CCL)
    • Replaced <memory_update> tags with -- memory blocks
    • Cleaner syntax, better readability, no XML escaping issues
    • Supports multi-line values via indentation
  • ResponseMonitor extracts and saves memory updates to SQLite
  • Updated documentation (README.md, GUIDE.md, SKILL.md files)

Removed

  • /loop <prompt> command - no longer needed
  • /meta_loop <prompt> command - replaced by automatic memory
  • /metamem command - functionality merged into /recall
  • ralph-loop dependency - was never implemented

Fixed

  • _save_to_memory() method - Fixed bug where memory update saving depended on recent_messages entry
    • Now saves memory_update independently, even if conversation not from Telegram
  • find_latest_transcript() - Fixed transcript file path detection
    • Now searches both ~/.claude/transcripts/ and ~/.claude/projects/*/ directories
    • Correctly finds Claude Code transcript files in project subdirectories
  • clear_all() and delete() methods - Fixed FTS5 index deletion order bug
    • FTS5 records must be deleted BEFORE main table records (due to content_rowid constraint)
    • Previously deletion order was incorrect, causing FTS5 index to become out of sync

Added

  • Self-referential meta-instruction storage - Auto-memory instruction now stored in SQLite DB
  • _get_or_init_auto_memory_instruction() method loads instruction from DB at startup
  • Default instruction auto-initialized to DB on first run
  • Instruction can be updated via /remember or auto-memory for iterative optimization
  • Automatic memory extraction on every Telegram message

[1.0.0] - 2025-01-31

Core Features

Telegram Bot Bridge

  • Polling-based message handling - Real-time message reception via Telegram Bot API
  • Command system - 10 built-in commands for session management
  • Response monitoring - Automatic detection and forwarding of Claude responses
  • Typing indicators - Real-time feedback during processing

Session Management

Command Description
/clear Clear current conversation
/continue_ Continue most recent session
/resume Interactive session picker with inline keyboard
/stop Interrupt Claude (sends Escape)
/status Check tmux and bridge status

Local Memory System

  • Fully automatic - Every message auto-extracts key info to memory (no command needed)
  • SQLite-based storage - Local database at ~/.matecode/memory.db
  • FTS5 full-text search - Fast relevance-ranked memory retrieval
  • Auto-save conversations - Every Q&A pair saved automatically
  • Memory injection - Relevant memories auto-injected into context
  • Manual management - /remember, /recall, /forget commands
  • Meta-update tracking - Self-referential memories auto-saved via -- memory blocks
  • Self-referential instruction - Meta-instruction stored in DB, iteratively improvable

tmux Integration

  • Session management - Automatic tmux session creation and attachment
  • Claude Code wrapper - Full tmux control via Telegram
  • Multi-window support - Switch between Claude and shell
  • Automatic attachment - matecode.sh auto-attaches after start/restart

Skills System

25+ integrated skills for extended functionality:

Category Skills
Development coding-agent, skill-creator, vercel-react-best-practices, remotion-best-practices
AI/ML openai-image-gen, openai-whisper, gemini, model-usage
Productivity notion, obsidian, slack, discord, github
Media video-frames, gifgrep, nano-pdf, summarize
Web agent-browser, canvas, web-design-guidelines
Utilities tmux, session-logs, blogwatcher, clawdhub
MateCode Self-management skill

Architecture

┌─────────────────┐     ┌──────────────┐     ┌─────────────────┐
│  Telegram Bot   │◄───►│  bridge.py   │◄───►│  Claude Code    │
│  (User Interface)│     │  (Bridge)    │     │  (tmux session) │
└─────────────────┘     └──────────────┘     └─────────────────┘
                               │
                               ▼
                        ┌──────────────┐
                        │  memory.py   │
                        │  (SQLite)    │
                        └──────────────┘

File Structure

MateBot/
├── bridge.py              # Main bridge implementation
├── memory.py              # Local memory system
├── matecode.sh            # Management script (start/stop/status/attach)
├── bridge_manager.sh      # Process management
├── README.md              # English documentation
├── GUIDE.md               # Chinese guide
├── TMUX_SETUP.md          # tmux configuration guide
├── .tmux.conf.local       # Custom tmux config
├── hooks/
│   └── send-to-telegram.sh  # Response hook script
├── matecode/
│   └── SKILL.md           # Skill metadata
└── skills/                # 25+ integrated skills

Environment Variables

Variable Default Description
TELEGRAM_BOT_TOKEN - Bot token (required)
TMUX_SESSION claude tmux session name
MEMORY_ENABLED true Enable memory system
MEMORY_MAX_RESULTS 5 Max memories per query
MEMORY_MAX_CONTEXT 2000 Context injection limit

Technical Highlights

  • Pure Python 3 - No external dependencies beyond standard library
  • Thread-safe - Concurrent response monitoring and message handling
  • JSONL transcript parsing - Efficient Claude output extraction
  • Hook-based integration - Non-intrusive Claude Code extension
  • Polling architecture - Reliable message delivery without webhooks
  • Automatic memory - Every prompt includes auto-memory instruction, responses parsed for -- memory blocks

Auto-Memory Flow

Startup → _get_or_init_auto_memory_instruction()
                    ↓
         [Load from DB or Initialize Default]
                    ↓
User Message → Bridge → _build_full_prompt()
                    ↓
         [DB-stored instruction prepended]
                    ↓
              Claude Response
                    ↓
         [Contains -- memory block]
                    ↓
         ResponseMonitor._save_to_memory()
                    ↓
              SQLite memory.db
                    ↓
    [Instruction itself can be updated via memory]

Future Roadmap

  • Webhook mode for lower latency
  • Multi-user support with access control
  • Conversation threading
  • File upload/download support
  • Voice message support
  • Custom skill marketplace integration