claw-compactor（爪式压实机）

name: claw-compactor version: "1.0.0" description: > Claw Compactor - 6-layer token compression skill for OpenClaw agents. Cuts workspace token spend by 50–97% using deterministic rule-engines plus Engram: a real-time, LLM-driven Observational Memory system. Run at session start for automatic savings reporting.

triggers: - "compress memory", "compress workspace", "save tokens", "token savings", "compress context", "run engram", "engram observe", "engram reflect", "memory compression", "benchmark compression"

Claw Compactor - OpenClaw Skill Reference

Overview

Claw Compactor reduces token usage across the full OpenClaw workspace using 6 compression layers:

1, Name=Rule Engine, Cost=Free, Notes=Dedup, strip filler, merge sections 2, Name=Dictionary Encoding, Cost=Free, Notes=Auto-codebook, $XX substitution 3, Name=Observation Compression, Cost=Free, Notes=Session JSONL → structured summaries 4, Name=RLE Patterns, Cost=Free, Notes=Path/IP/enum shorthand 5, Name=Compressed Context Protocol, Cost=Free, Notes=Format abbreviations 6, Name=Engram, Cost=LLM API, Notes=Real-time Observational Memory

Skill location: skills/claw-compactor/ Entry point: scripts/mem_compress.py Engram CLI: scripts/engram_cli.py

Auto Mode (Recommended - Run at Session Start)

python3 skills/claw-compactor/scripts/mem_compress.py <workspace> auto

Automatically compresses all workspace files, tracks token counts between runs, and reports savings. Run this at the start of every session.

Core Commands

Full Pipeline (All Layers)

python3 scripts/mem_compress.py <workspace> full Runs all 5 deterministic layers in optimal order. Typical: 50%+ combined savings.

Benchmark (Non-Destructive)

python3 scripts/mem_compress.py <workspace> benchmark

JSON output:

python3 scripts/mem_compress.py <workspace> benchmark --json Dry-run report showing potential savings without writing any files.

Layer 1: Rule-based compression

python3 scripts/mem_compress.py <workspace> compress

Layer 2: Dictionary encoding

python3 scripts/mem_compress.py <workspace> dict

Layer 3: Observation compression (session JSONL → summaries)

python3 scripts/mem_compress.py <workspace> observe

Layer 5: Tokenizer optimization

python3 scripts/mem_compress.py <workspace> optimize

Tiered summaries (L0/L1/L2)

python3 scripts/mem_compress.py <workspace> tiers

Cross-file deduplication

python3 scripts/mem_compress.py <workspace> dedup

Token count report

python3 scripts/mem_compress.py <workspace> estimate

Workspace health check

python3 scripts/mem_compress.py <workspace> audit

Global Options

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

Engram - Layer 6: Real-Time Observational Memory

Engram is the flagship layer. It operates as a live engine alongside conversations, automatically compressing messages into structured, priority-annotated knowledge.

Prerequisites

Configure via engram.yaml (recommended) or environment variables:

# engram.yaml - place in claw-compactor root
llm:
 provider: openai-compatible
 base_url: http://localhost:8403
 model: claude-code/sonnet
 max_tokens: 4096

threads:
 default:
 observer_threshold: 30000 # pending tokens before Observer fires
 reflector_threshold: 40000 # observation tokens before Reflector fires

concurrency:
 max_workers: 4 # parallel thread workers

# Alternative: environment variables
export ANTHROPIC_API_KEY=sk-ant-... # Preferred

# or
export OPENAI_API_KEY=sk-... # OpenAI-compatible fallback
export OPENAI_BASE_URL=https://... # Optional: custom endpoint (local LLM, etc.)

### Engram Auto-Mode (Recommended for Production)
Auto-detects all active threads and processes them concurrently (4 workers):

# Single run - auto-detects all threads
python3 scripts/engram_auto.py --workspace ~/.openclaw/workspace

# Via shell wrapper
bash scripts/engram-auto.sh

# Via CLI
python3 scripts/engram_cli.py <workspace> auto --config engram.yaml
python3 scripts/engram_cli.py <workspace> status --thread openclaw-main
python3 scripts/engram_cli.py <workspace> observe --thread openclaw-main
python3 scripts/engram_cli.py <workspace> reflect --thread openclaw-main

**Retry:** LLM calls retry on 429/5xx with exponential backoff (2s→4s→8s, max 3 attempts).
No retry on 400/401/403 (fail fast on config errors).

# Check all thread statuses
python3 scripts/mem_compress.py <workspace> engram status

# Force Observer for a thread
python3 scripts/mem_compress.py <workspace> engram observe --thread <thread-id>

# Force Reflector for a thread
python3 scripts/mem_compress.py <workspace> engram reflect --thread <thread-id>

# Print injectable context
python3 scripts/mem_compress.py <workspace> engram context --thread <thread-id>

# Status: all threads
python3 scripts/engram_cli.py <workspace> status

# Status: single thread
python3 scripts/engram_cli.py <workspace> status --thread <thread-id>

# Force observe
python3 scripts/engram_cli.py <workspace> observe --thread <thread-id>

# Force reflect
python3 scripts/engram_cli.py <workspace> reflect --thread <thread-id>

# Import conversation from file (JSON array or JSONL)
python3 scripts/engram_cli.py <workspace> ingest \
 --thread <thread-id> --input /path/to/conversation.jsonl

# Get injectable context string (ready for system prompt)
python3 scripts/engram_cli.py <workspace> context --thread <thread-id>

# JSON output for any command
python3 scripts/engram_cli.py <workspace> status --json
python3 scripts/engram_cli.py <workspace> context --thread <id> --json

# Start daemon, pipe JSONL messages via stdin
python3 scripts/engram_cli.py <workspace> daemon --thread <thread-id>

# Pipe a message:
echo '{"role":"user","content":"Hello!","timestamp":"12:00"}' | \

# Control commands (send as JSONL):
echo '{"__cmd":"observe"}' # force observe now
echo '{"__cmd":"reflect"}' # force reflect now
echo '{"__cmd":"status"}' # print thread status JSON
echo '{"__cmd":"quit"}' # exit daemon

# Quiet mode (suppress startup messages on stderr)
python3 scripts/engram_cli.py <workspace> daemon --thread <id> --quiet

### Engram Python API

from scripts.lib.engram import EngramEngine

engine = EngramEngine( workspace_path="/path/to/workspace", observer_threshold=30_000, # tokens before auto-observe reflector_threshold=40_000, # tokens before auto-reflect anthropic_api_key="sk-ant-...", # or set ANTHROPIC_API_KEY env )

Add a message - auto-triggers observe/reflect when thresholds exceeded

status = engine.add_message("thread-id", role="user", content="Hello!")

Manual trigger regardless of thresholds

obs_text = engine.observe("thread-id") # returns None if no pending msgs ref_text = engine.reflect("thread-id") # returns None if no observations

Get full context dict

ctx = engine.get_context("thread-id")

Build injectable system context string

ctx_str = engine.build_system_context("thread-id")

Ready to prepend to system prompt

Engram Configuration Variables

ANTHROPIC_API_KEY, Default=-, Description=Anthropic API key (preferred) OPENAI_API_KEY, Default=-, Description=OpenAI-compatible API key OPENAI_BASE_URL, Default=https://api.openai.com, Description=Custom endpoint for local LLMs OM_OBSERVER_THRESHOLD, Default=30000, Description=Pending tokens before auto-observe OM_REFLECTOR_THRESHOLD, Default=40000, Description=Observation tokens before auto-reflect OM_MODEL, Default=claude-opus-4-5, Description=LLM model override

Threshold Tuning Quick Reference

Each Observer call ≈ 2K output tokens (Sonnet). Daily volume at default 30K threshold:

#aimm, Daily Tokens=~149K, @30K threshold=~5×/day, @10K threshold=~15×/day openclaw-main, Daily Tokens=~138K, @30K threshold=~4.5×/day, @10K threshold=~14×/day #open-compress, Daily Tokens=~68K, @30K threshold=~2.3×/day, @10K threshold=~7×/day #general, Daily Tokens=~62K, @30K threshold=~2×/day, @10K threshold=~6×/day subagent, Daily Tokens=~43K, @30K threshold=~1.4×/day, @10K threshold=~4×/day cron, Daily Tokens=~9K, @30K threshold=~0.3×/day, @10K threshold=~1×/day Total, Daily Tokens=~470K/day, @30K threshold=~16×/day (~32K output tokens), @10K threshold=~47×/day (~94K output tokens)

Start at observer_threshold: 30000. Tune down for fresher context; tune up to reduce cost.

Engram Benchmark Summary

| Engram (L6) | 87.5% | 0.038 | 0.414 | ~35s | 2 | | RuleCompressor (L1–5) | 9.0% | 0.923 | 0.958 | ~6ms | 0 | | RandomDrop | 21.5% | 0.852 | 0.911 | ~0ms | 0 |

• Engram low ROUGE-L = semantic restructuring, not verbatim copy - intent is preserved
• Use RuleCompressor for instant prompt compression; Engram for long-term memory
• Full results → benchmark/RESULTS.md

Observation Format

Engram produces structured, bilingual (EN/中文) priority-annotated logs:

Date: 2026-03-05

• 12:10 User building OpenCompress; deadline one week / 用户在构建 OpenCompress,deadline 一周内

- 12:10 Using ModernBERT-large / 使用 ModernBERT-large - 🟡 12:12 Discussed annotation strategy / 讨论了标注策略

• 🟡 12:30 Deployment pipeline discussion on M3 Ultra
• 🟢 12:45 User prefers concise replies

• Critical - goals, deadlines, blockers, key decisions (never dropped)
• 🟡 Important - technical details, ongoing work, preferences
• 🟢 Useful - background, mentions, soft context

Memory Storage Layout

memory/engram/{thread_id}/ ├── pending.jsonl # Unobserved message buffer (auto-cleared after observe) ├── observations.md # Observer output - append-only structured log ├── reflections.md # Reflector output - compressed long-term memory (overwrites) └── meta.json # Timestamps and token counts

Integration with OpenClaw Memory System

System Prompt Injection

Inject Engram context at the start of each session:

engine = EngramEngine(workspace_path) ctx_str = engine.build_system_context("my-session") if ctx_str: system_prompt = ctx_str + "\ \ " + base_system_prompt

The build_system_context() output structure:

Long-Term Memory (Reflections)

<Reflector output - long-term compressed context>

Recent Observations

<Last 200 lines of Observer output>

<!-- engram_tokens: 1234 -->

Combining Engram with Deterministic Layers

After an Engram session, run the deterministic pipeline on the output files:

Then apply deterministic compression to further reduce those:

Recommended Workflow for Long-Running Agent Sessions

1. Session start: inject build_system_context() into system prompt
2. Each message: call engine.add_message() - auto-triggers observe/reflect
3. Session end / weekly cron: run full pipeline on workspace
4. Multi-session continuity: context persists in memory/engram/{thread}/

OpenClaw Skill Installation

To install as an OpenClaw skill, ensure the skill directory is available at: ~/.openclaw/workspace/skills/claw-compactor/ or configure the path in your OpenClaw skill registry.

SKILL.md is read by the OpenClaw agent dispatcher. The description and triggers fields above control when this skill is automatically activated.

Heartbeat / Cron Automation

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

Cron (Sunday 3am):
0 3 * * 0 cd /path/to/skills/claw-compactor && \
 python3 scripts/mem_compress.py /path/to/workspace full

## Output Artifacts Reference
Dictionary codebook, Location=`memory/.codebook.json`, Description=Must travel with memory files
Observed session log, Location=`memory/.observed-sessions.json`, Description=Tracks processed transcripts
Layer 3 summaries, Location=`memory/observations/`, Description=Observation compression output
Engram observations, Location=`memory/engram/{thread}/observations.md`, Description=Live Observer log
Engram reflections, Location=`memory/engram/{thread}/reflections.md`, Description=Distilled long-term memory
Level 0 summary, Location=`memory/MEMORY-L0.md`, Description=~200 token ultra-compressed summary
Level 1 summary, Location=`memory/MEMORY-L1.md`, Description=~500 token compressed summary

## Troubleshooting
- `FileNotFoundError` on workspace: Point path to workspace root containing `memory/`
- Dictionary decompression fails: Check `memory/.codebook.json` is valid JSON
- Zero savings on `benchmark`: Workspace already optimized
- `observe` finds no transcripts: Check `sessions/` for `.jsonl` files
- Engram: "no API key configured": Set `ANTHROPIC_API_KEY` or `OPENAI_API_KEY`
- Engram Observer returns `None`: No pending messages for that thread
- Token counts seem wrong: Install tiktoken: `pip3 install tiktoken`