Conversation Saver

Automatically extract key facts from conversation history and persist to local memory files. Silent background operation with rule+LLM hybrid extraction.

Automatically extracts key facts from your conversations and persists them to the appropriate local memory files (WARM_MEMORY.md, MEMORY.md, ontology, USER.md). Works silently in the background without interrupting the user flow.

When to Use

• You want your agent to remember important details from conversations automatically
• You need hands-off fact extraction (no interactive questioning)
• You want facts categorized into the right places (preferences, schedules, person details)
• You prefer local-first persistence (no external dependencies)

How It Works

1. Extraction Pipeline

text

Raw Messages → Rule-Based Filter → LLM Extraction → Classification → Deduplication → Persistence

• Rule-Based: Fast keyword/regex matching for obvious facts (dates, locations, names)
• LLM: Step 3.5 Flash extracts structured facts from ambiguous conversations
• Classification: Routes facts to the correct storage target
• Deduplication: Avoids recording the same fact multiple times

2. Storage Targets

Fact Type	Destination	Example
Person details (family, friends)	ontology + WARM_MEMORY.md (家庭)	"老婆去上海出差"
Time commitments	WARM_MEMORY.md (日程)	"下周二回来"
Locations	USER.md + WARM_MEMORY.md	"常驻武汉"
Preferences	WARM_MEMORY.md (互动偏好)	"不要一股脑发照片"
System rules	TOOLS.md / AGENTS.md	"回复必须@老布"
Important decisions	MEMORY.md	"决定用Tailwind"

3. Trigger Modes

• On Session End: Automatically run after each conversation (requires AGENTS.md hook)
• Heartbeat Backfill: Scan recent days for missed conversations (configurable)
• Manual: uv run scripts/extract.py --session <sessionKey> or --days N

Installation

bash

# From ClawHub (recommended)
clawhub install conversation-saver

# Or manual
cd ~/.openclaw/workspace/skills
git clone <your-repo> conversation-saver

Configuration

Edit config.json to customize behavior:

json

{
  "extraction": {
    "enabled": true,
    "auto_on_session_end": true,
    "heartbeat_reprocess_days": 2,
    "min_confidence": 0.6,
    "max_facts_per_session": 10
  },
  "filters": {
    "user_id": "ou_39f0f10fb55c7c782610cad6a97f4842",
    "ignore_bot_messages": true,
    "ignore_short_messages": true,
    "min_message_length": 5
  },
  "persistence": {
    "verify_after_write": true,
    "backup_before_write": false,
    "deduplicate_across_sessions": true
  }
}

Usage

Automatic Mode (Recommended)

Add to AGENTS.md to run after each session:

markdown

## Session End Hooks
- 每次会话结束 → 运行 conversation-saver

Or integrate into your existing heartbeat:

bash

uv run ~/.openclaw/workspace/skills/conversation-saver/scripts/extract.py --days 2 --reprocess

Manual Run

bash

# Extract from specific session
uv run scripts/extract.py --session <sessionKey>

# Backfill last 3 days
uv run scripts/extract.py --days 3 --reprocess

# Dry run (show what would be extracted)
uv run scripts/extract.py --session <sessionKey> --dry-run

Files Structure

text

conversation-saver/
├── SKILL.md
├── config.json
├── scripts/
│   ├── __init__.py
│   ├── extract.py      # Main entry point
│   ├── classifier.py   # Fact classification
│   ├── persister.py    # File writing with verification
│   └── utils.py        # Helpers (dedupe, date parsing)
└── README.md

Customization

Add Keywords

Edit scripts/extract.py → KEYWORD_CATEGORIES:

python

KEYWORD_CATEGORIES = {
    "person": ["老婆", "小美美", "包子", "家人", "朋友"],
    "location": ["武汉", "上海", "郑州", "出差", "旅行"],
    "time": ["周三", "周二", "本周", "下周", "月", "日"],
    "preference": ["喜欢", "不要", "记住", "记得", "规则"]
}

Adjust LLM Prompt

Modify scripts/extract.py → LLM_EXTRACTION_PROMPT to change extraction style or output format.

Requirements

• Python 3.10+
• OpenClaw agent with tool access (read, write, edit)
• StepFun model (for LLM extraction)

Testing

bash

# Dry run on today's conversations
uv run scripts/extract.py --today --dry-run

# Check extracted facts count
uv run scripts/extract.py --today --stats-only

Limitations

• Single-user focus: Designed for one primary user (your USER.md)
• No vector search: Facts are stored in files, not semantic searchable (yet)
• Language: Optimized for Chinese conversations (keywords in Chinese)
• No GUI: All configuration via config.json

Credits

Inspired by openclaw-user-profiler's structured approach and elite-longterm-memory's tiered architecture.