Claw Compactor

Claw Compactor — OpenClaw Skill Reference

Overview

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

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.

Individual Layers

# 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 4: RLE pattern encoding (runs inside `compress`)
# 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).

Engram via Unified Entry Point

# 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>

Engram via Dedicated CLI

# 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

Engram Daemon Mode (Real-Time Streaming)

# 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"}' | \
    python3 scripts/engram_cli.py <workspace> daemon --thread <thread-id>

# 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!")
# Returns: {"observed": bool, "reflected": bool, "pending_tokens": int, ...}

# 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")
# Returns: {"thread_id", "observations", "reflection", "recent_messages", "stats", "meta"}

# Build injectable system context string
ctx_str = engine.build_system_context("thread-id")
# Ready to prepend to system prompt

Engram Configuration Variables

Threshold Tuning Quick Reference

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

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

Engram Benchmark Summary

• 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:

from scripts.lib.engram import EngramEngine

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

# Engram produces observations.md and reflections.md
# Then apply deterministic compression to further reduce those:
python3 scripts/mem_compress.py <workspace> full

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

Troubleshooting