cut-your-tokens-97percent-savings-on-session-transcripts-via-observation-extraction（通过观察提取，在会话记录上节省 97% 的代币）

🦞 Claw Compactor

Claw Compactor Banner

*"Cut your tokens. Keep your facts."*

Cut your AI agent's token spend in half. One command compresses your entire workspace — memory files, session transcripts, sub-agent context — using 5 layered compression techniques. Deterministic. Mostly lossless. No LLM required.

Features

• 5 compression layers working in sequence for maximum savings
• Zero LLM cost — all compression is rule-based and deterministic
• Lossless roundtrip for dictionary, RLE, and rule-based compression
• ~97% savings on session transcripts via observation extraction
• Tiered summaries (L0/L1/L2) for progressive context loading
• CJK-aware — full Chinese/Japanese/Korean support
• One command (full) runs everything in optimal order

5 Compression Layers

\*Lossy techniques preserve all facts and decisions; only verbose formatting is removed.

Quick Start

git clone https://github.com/aeromomo/claw-compactor.git
cd claw-compactor

# See how much you'd save (non-destructive)
python3 scripts/mem_compress.py /path/to/workspace benchmark

# Compress everything
python3 scripts/mem_compress.py /path/to/workspace full

Requirements: Python 3.9+. Optional: pip install tiktoken for exact token counts (falls back to heuristic).

Architecture

┌─────────────────────────────────────────────────────────────┐
│                      mem_compress.py                        │
│                   (unified entry point)                     │
└──────┬──────┬──────┬──────┬──────┬──────┬──────┬──────┬────┘
       │      │      │      │      │      │      │      │
       ▼      ▼      ▼      ▼      ▼      ▼      ▼      ▼
  estimate compress  dict  dedup observe tiers  audit optimize
       └──────┴──────┴──┬───┴──────┴──────┴──────┴──────┘
                        ▼
                  ┌────────────────┐
                  │     lib/       │
                  │ tokens.py      │ ← tiktoken or heuristic
                  │ markdown.py    │ ← section parsing
                  │ dedup.py       │ ← shingle hashing
                  │ dictionary.py  │ ← codebook compression
                  │ rle.py         │ ← path/IP/enum encoding
                  │ tokenizer_     │
                  │   optimizer.py │ ← format optimization
                  │ config.py      │ ← JSON config
                  │ exceptions.py  │ ← error types
                  └────────────────┘

Commands

All commands: python3 scripts/mem_compress.py <workspace> <command> [options]

Global Options

• --json — Machine-readable JSON output
• --dry-run — Preview changes without writing
• --since YYYY-MM-DD — Filter sessions by date
• --auto-merge — Auto-merge duplicates (dedup)

Real-World Savings

cacheRetention — Complementary Optimization

Before compression runs, enable prompt caching for a 90% discount on cached tokens:

{
  "models": {
    "model-name": {
      "cacheRetention": "long"
    }
  }
}

Compression reduces token count, caching reduces cost-per-token. Together: 50% compression + 90% cache discount = 95% effective cost reduction.

Heartbeat Automation

Run weekly or on heartbeat:

## Memory Maintenance (weekly)
- python3 skills/claw-compactor/scripts/mem_compress.py <workspace> benchmark
- If savings > 5%: run full pipeline
- If pending transcripts: run observe

Cron example:

0 3 * * 0 cd /path/to/skills/claw-compactor && python3 scripts/mem_compress.py /path/to/workspace full

Configuration

Optional claw-compactor-config.json in workspace root:

{
  "chars_per_token": 4,
  "level0_max_tokens": 200,
  "level1_max_tokens": 500,
  "dedup_similarity_threshold": 0.6,
  "dedup_shingle_size": 3
}

All fields optional — sensible defaults are used when absent.

Artifacts

FAQ

Q: Will compression lose my data? A: Rule engine, dictionary, RLE, and tokenizer optimization are fully lossless. Observation compression and CCP are lossy but preserve all facts and decisions.

Q: How does dictionary decompression work? A: decompress_text(text, codebook) expands all $XX codes back. The codebook JSON must be present.

Q: Can I run individual steps? A: Yes. Every command is independent: compress, dict, observe, tiers, dedup, optimize.

Q: What if tiktoken isn't installed? A: Falls back to a CJK-aware heuristic (chars÷4). Results are ~90% accurate.

Q: Does it handle Chinese/Japanese/Unicode? A: Yes. Full CJK support including character-aware token estimation and Chinese punctuation normalization.

Troubleshooting

• FileNotFoundError on workspace: Ensure path points to workspace root (contains memory/ or MEMORY.md)
• Dictionary decompression fails: Check memory/.codebook.json exists and is valid JSON
• Zero savings on benchmark: Workspace is already optimized — nothing to do
• observe finds no transcripts: Check sessions directory for .jsonl files
• Token count seems wrong: Install tiktoken: pip3 install tiktoken

Credits

• Inspired by claude-mem by thedotmack
• Built by Bot777 🤖 for OpenClaw

License

MIT