🦞 灵枢 AutoBrain

Give your OpenClaw agent a brain — long-term memory, anti-hallucination, self-evolution, and workflow orchestration, out of the box.

---

✨ What It Does

灵枢 AutoBrain is a plugin + skill hybrid pack that upgrades any OpenClaw agent from a stateless chatbot to a persistent, self-improving AI companion. It hooks into the agent lifecycle and deploys 30+ Python engines across 8 groups — no manual setup required.

Capability	What You Get
🧠 Long-term Memory	5-layer memory (session → short-term → dream consolidation → archive → vector retrieval). Remembers who you are between sessions.
🛡️ Anti-Hallucination	Authority domain whitelist, knowledge assertion validation, multi-source cross-verification (16 engines).
🔄 Self-Evolution	LLM-as-Judge self-scoring + Reflexion reflection (failure pattern, root cause, fix strategy).
🎯 Dual-Mode + R-CCAM	Fast path for simple Q&A, deep reasoning for complex tasks. R-CCAM classifier for query complexity routing.
🔮 MASA Prediction-Alignment	Pre-task difficulty/time/confidence estimation + post-task deviation analysis + feedback to self-evolution & parameter auto-tuning
✅ 8 Iron Rules	Prevents "I thought so" mistakes — fact-check before answering, verify before acting.
🩺 Health Check	Daily cron (00:00) — engine status, disk space, gateway health, anomaly detection.
🔌 Failover	Auto-switch to backup model when primary fails. Circuit breaker prevents cascade failures.
🔀 Workflow Orchestration	Multi-skill coordination, conflict detection, task routing, rule engine.
📋 Context Capsule (DAG)	DAG-based context management with incremental summarization. SQLite-backed, retains message dependencies.
🧠 Memory Consolidation	CLS complementarity learning + offline replay + interference merging + predictive conflict detection.
🔍 Enhanced Retrieval	Query rewriting + RRF fusion + retrieval confidence evaluation for higher recall accuracy.

🚀 Quick Start

Prerequisites

• OpenClaw ≥ 2026.5.0
• Node.js ≥ 18
• Python ≥ 3.10

Install

# Download the tarball, then:
# ⚠️ This package is a plugin-primary + skill metadata hybrid pack, exclusively occupying the OpenClaw plugin slot.
# Only one crusheart-* family plugin can be active at a time.

openclaw plugins install --dangerously-force-unsafe-install ./crusheart-pack/

# Restart gateway
supervisorctl restart openclaw-gateway
# or: openclaw gateway restart

Why --dangerously-force-unsafe-install? The plugin's index.js calls child_process to execute Python engines during bootstrap. This is a false positive — the code is local, not remote. We'll migrate to pure SDK hooks in a future release.

Plugin Slot (Exclusive)

Exclusive slot locking implemented via .crusheart-slot.json. Only one crusheart-family plugin can be active at a time. To switch versions, uninstall the old plugin first, then install the new plugin.

Compat Layer

Built-in core/engines/compat/ compatibility engine group (3 modules: compat_engine, compat_registry, compat_scaffold) supporting third-party engine hot-swapping and registration integration.

One-time Run Items on First Install (Non-cron Tasks)

• Memory Auto-Scan: scan_memory.py — Scans memory/ directory on first install, automatically ingests extracted memories into the memory system
• Skill Classification Index: scan_skills.py — Scans skills/ directory on first install to build classification index
• Self-Correction Initialization: init_correction_data.py — Establishes JudgeEngine data pipeline
• Version Check: version_check.py — Checks for new versions on clawhub.ai and notifies when updates are available (once only)

Verify

# Check plugin is loaded
openclaw plugins list | grep crusheart

# Watch the bootstrap log — you should see:
#   ✅ 31 engines initialized
#   ✅ Pipeline: 10 stages ready
#   ✅ Daily maintenance cron registered

🏗️ Architecture

openclaw-gateway
  │
  ├── crusheart-autobrain-turbo (plugin)
  │     ├── index.js                    ← Node.js bridge to Python engines
  │     ├── bundle/crusheart-core.tar.gz  ← 30+ Python engines (deployed at bootstrap)
  │     └── skill/                       ← OpenClaw skill metadata
  │
  ├── core/engines/                     ← Deployed to workspace
  │     ├── init/      (11)  — Auto-loader, config validator, session bootstrap
  │     ├── memory/     (6)  — 5-layer memory, vector index, user portrait
  │     ├── quality/   (10)  — Anti-hallucination, judge engine, circuit breaker
  │     ├── operations/ (7)  — Health check, decision core, background scheduler
  │     ├── workflow/   (7)  — Orchestrator, rule engine, serial lanes
  │     ├── tools/     (12)  — Failover, DB, template library, plugin SDK
  │     ├── hooks/      (4)  — Dual-mode classifier, self-evolution
  │     └── compat/     (3)  — Third-party engine hot-plug
  │
  └── core/pipeline/                    ← 10-stage message pipeline
        ├── stage 0: Engine status probe
        ├── stage 1: Dual-mode classification
        ├── stage 2: Skill matching
        ├── stage 3: Anti-hallucination check
        ├── stage 4: Engine routing
        ├── stage 5: Session state (hot RAM layer)
        ├── stage 6: Memory alignment
        ├── stage 7: Evolution context injection
        ├── stage 8: Self-reflection
        └── stage 9: Outbound anti-hallucination

🧠 Memory System Deep Dive

The memory system is the crown jewel. Five layers, one unified API:

Session (L1) ──► Short-term (L2) ──► Dream (L3) ──► Archive (L4) ──► Vector (L5)
    │                 │                   │               │
    ▼                 ▼                   ▼               ▼
  Ephemeral      7-day TTL         Read-only,       Permanent,
  (current        consolidated      consolidated     never deleted
   conversation)  by sleep cron     weekly

Key features:

• Inverted index for keyword search
• TF-IDF + cosine similarity for semantic reranking
• Cold/hot storage — frequently accessed memories stay hot, old ones decay gracefully
• Scene grouping — memories tagged by context (work, chat, health, etc.)
• Anti-forget engine — periodic review prevents important memories from decaying
• Global mutex with heartbeat — prevents duplicate cron execution

📦 What's Included

Artifact	Path	Purpose
Plugin entry	index.js	Bridges Python engines into OpenClaw lifecycle hooks
Plugin manifest	openclaw.plugin.json	OpenClaw plugin registration
Engine bundle	bundle/crusheart-core.tar.gz	30+ Python engines + 10-stage pipeline
Skill metadata	skill/_meta.json	Skill marketplace registration
Skill doc	skill/SKILL.md	Skill documentation
Architecture doc	bundle/ARCHITECTURE.md	Full system architecture reference
Auto-deployed scripts	bundle/daily_maintenance.py, bundle/version_check.py, bundle/scan_memory.py, bundle/scan_skills.py, bundle/read_config.py, bundle/init_correction_data.py, bundle/register_crons.sh	Plugin auto-deploys to workspace scripts/ at bootstrap
Cron tasks	(auto-configured via hook)	Daily maintenance (01:00), engine init (05:00)

🔧 Configuration

Environment Variables (optional,requires you to configure your own API key)

Variable	Purpose	Default
EMBEDDING_API_URL	Remote embedding service URL	(local TF-IDF fallback)
EMBEDDING_API_KEY	Bearer token for embedding API	(none)

Without these, the system falls back to local TF-IDF vector search — slightly less accurate but fully functional.

Cron Jobs (auto-registered)

Time	Name	Description
0 1 * * *	Daily Task Summary & System Cleanup	Unified task: health inspection + garbage cleanup + self-correction maintenance + memory maintenance + Dream scan + Replay distillation + execution review
0 5 * * *	Engine Init + Version Check	Reload engine modules; check for new versions on clawhub.ai on install (first install only)

Memory scan archival, skill classification scan, and self-correction chain initialization run once on first install only, not as scheduled tasks.

🤝 Contributing

This is an early-stage open-source project. Contributions welcome!

• Issues: Bug reports, feature requests
• PRs: Engine improvements, new pipeline stages, documentation
• Discussions: Architecture questions, use cases

Development

# After cloning, manually deploy to workspace:
tar xzf bundle/crusheart-core.tar.gz -C /path/to/workspace
# Edit Python engines in core/engines/
# Edit plugin bridge in index.js
# Repack:
tar czf bundle/crusheart-core.tar.gz -C /tmp/repack .

📄 License

MIT0 — see LICENSE.

The current version is under development and testing. If you have any suggestions or questions about this plugin, please send them to the email HIM603070@gmail.com