🧠 Brainplex

The intelligence layer for AI agents.

npx brainplex init

One command. 60 seconds. Your agents get trust scoring, runtime policies, persistent memory, self-learning from mistakes, and health monitoring.

Before and After

Before Brainplex:

• Your agent has full permissions — day and night, in production and dev
• Every session starts from zero — no memory of yesterday
• You find out something broke when a user complains
• Sub-agents inherit all parent permissions without boundaries

After Brainplex:

• Every agent has a trust score. Permissions scale with behavior. Night mode blocks risky ops at 3am.
• Agents remember conversations, decisions, and commitments across sessions
• Anomaly detection catches problems before users do
• Cross-agent trust ceilings prevent permission escalation

Real example: You install Brainplex, go to sleep. Next morning:

⚡ NOTABLE (while you were away)
· 🚫 forge tried exec("rm -rf /tmp/cache") at 03:22 — DENIED (night mode)
· 🔒 2 API keys redacted from agent output (credential guard)
· main made 47 governed calls between 00:00–06:00 — all clean

Install

npx brainplex init

Brainplex detects your agents, installs the modules, generates configs, and wires everything in:

🧠 Brainplex — The Intelligence Layer for AI Agents

🔍 Scanning...
   ✓ Found 3 agents: main, forge, cerberus

📦 Installing...
   ✓ Governance · Cortex · Membrane · Leuko

⚙️  Configuring...
   ✓ Trust: main=60, forge=45, cerberus=50, *=10
   ✓ Night mode: 23:00–06:00
   ✓ Credential guard: ON

✓ Done — run: openclaw gateway restart

| Flag | Effect | |------|--------| | --full | Include Knowledge Engine (entity extraction, fact store) | | --dry-run | Preview without making changes | | --config <path> | Custom openclaw.json path | | --verbose | Show npm install output |

Dashboard

After installing, type /brainplex in any agent conversation:

Trust scores, governance stats, notable events, shield score — one command.

The Five Modules

🔒 Governance — The Firewall

Controls what agents can do and when.

• Trust scoring — per-agent scores that rise with clean behavior, drop with violations
• Policy engine — YAML-based rules with conditions (time, tool, context, trust level)
• Night mode — block destructive commands outside business hours
• Credential guard — 3-layer detection (regex → entropy → hash vault), redaction before output
• Approval workflows — require human sign-off for sensitive operations
• Audit trail — append-only JSONL with ISO 27001 controls
• Cross-agent governance — parent policies propagate to sub-agents, trust ceilings prevent escalation

🧠 Cortex — The Brain

Learns from what your agents do.

• Thread tracking — auto-detect open work items from conversations
• Decision tracking — extract and persist decisions with impact classification
• Commitment tracking — detect promises, track fulfillment, flag overdue items
• Trace analysis — 3-stage pipeline: detect → classify (LLM) → generate rules
• 7 signal detectors — corrections, doom loops, hallucinations, repeat failures, dissatisfaction, unverified claims, tool failures
• Auto-generated rules — findings become SOUL.md rules, governance policies, or cortex patterns
• 10-language support — pattern detection in EN, DE, FR, ES, PT, IT, JA, KO, RU, ZH

💾 Membrane — The Memory

Gives agents persistent context across sessions.

• Episodic memory — what happened in past conversations
• Semantic memory — facts and relationships extracted from interactions
• Working memory — current task context and state
• Agent isolation — each agent's memories are separated, no cross-contamination
• Salience filtering — retrieves only what's relevant to the current conversation

⚕️ Leuko — The Immune System

Monitors agent health and catches problems early.

• Anomaly detection — directory growth spikes, declining metrics, trend analysis
• Bootstrap integrity — checks that critical config files haven't drifted
• Pipeline correlation — detects when failures in one system affect another
• LLM-powered recommendations — actionable suggestions based on current system state
• Goal quality checks — flags stale or poorly defined goals

📚 Knowledge Engine — The Knowledge Base (optional)

Structured knowledge extraction and retrieval.

• Entity extraction — regex + LLM-powered extraction of people, orgs, products, dates
• Fact store — SPO triples (Subject-Predicate-Object) with relevance decay
• Vector embeddings — ChromaDB sync for semantic search
• Deduplication — automatic merging of duplicate entities and facts

Install with --full flag.

How They Work Together

Message in → Governance (policy + trust gate)
          → Membrane (inject relevant memories)
          → Agent processes
          → Governance (response gate + credential redaction + audit)
          → Cortex (track threads, decisions; analyze traces)
          → Leuko (health monitoring)
          → Knowledge Engine (extract entities, store facts)

Each module works independently — install just one if that's all you need. But together, they compound: Cortex learns from Governance denials. Leuko monitors Membrane health. Knowledge Engine enriches Cortex's context.

Design

• Zero telemetry — no tracking, no analytics, no phone-home
• Zero dependencies — Node.js builtins only
• Never overwrites existing configs (backs up first)
• Never auto-restarts your gateway
• Idempotent — run it twice, nothing breaks
• MIT licensed — fork it, modify it, sell it

Requirements

• Node.js ≥ 22
• OpenClaw with openclaw.json

Individual Modules

Don't want the full suite? Each module is independently installable:

npm install @vainplex/openclaw-governance    # Policy engine, trust, approval, redaction
npm install @vainplex/openclaw-cortex        # Conversation intelligence, trace analysis
npm install @vainplex/openclaw-membrane      # Episodic memory
npm install @vainplex/openclaw-leuko         # Health monitoring
npm install @vainplex/openclaw-knowledge-engine  # Entity extraction, fact store

Contributing

github.com/alberthild/vainplex-openclaw

License

MIT © Albert Hild