AI Starter Package

AI Brain Blueprint

The complete architecture of an AI brain system. Understand how every component connects and works together to create an autonomous AI workforce. See how it works →

System Architecture

Want to go deeper? Read the context engineering guide for how these layers interact at runtime.

ScreenpipeLayer 0: Foundation

Screen memory — captures everything you see and hear. Searchable AI memory for your entire digital life.

CLAUDE.mdLayer 1: Core Brain

Identity, rules, architecture decisions, conventions — your AI's operating manual

Model RoutingLayer 1: Core Brain

Multi-model support — route tasks to Haiku (fast), Sonnet (balanced), or Opus (deep) by complexity

250+ Specialist AgentsLayer 2: Agents

11 OpenClaw + 98 Ruflo swarm + 150+ Agency Agents personalities (engineering, design, sales, marketing, testing, game dev)

1,738 SkillsLayer 2: Agents

1,703 global + 35 project SKILL.md files — verified count, loaded on demand per task

Human-in-the-LoopLayer 2: Agents

Auditor-gated knowledge promotion, safety checks, approval gates

7-Tier Memory StackLayer 3: Memory

Session → Working → Knowledge Base → Agent Memory → Knowledge Graph → Daily Archive

RAG & RetrievalLayer 3: Memory

Knowledge graph (MCP), semantic search, retrieval-augmented generation

Ruflo Vector MemoryLayer 3: Memory

Hybrid backend — SQLite structured data + HNSW vector search for sub-millisecond semantic retrieval

MCP ServersLayer 4: Integration

330+ tool connections — GitHub, Slack, Apify, SerpAPI, Ruflo MCP (310+ tools), Browserbase, and more

Obsidian Second BrainLayer 4: Integration

Bidirectional sync — daily logs, decision records, tool reviews, linked knowledge vault

Ruflo Swarm EngineLayer 5: Orchestration

Multi-agent swarms — Queen/worker topology, consensus coordination, autopilot mode, anti-drift gates

Commands & HooksLayer 5: Orchestration

Slash commands (/start, /sync, /wrap-up) + 10 Ruflo command groups + 7 automated hook types

Self-Improvement LoopLayer 6: Evaluation

Observation → Nomination → Auditor Review → Knowledge Promotion (RETRIEVE → JUDGE → DISTILL)

Agent TrainingLayer 6: Evaluation

Agent Lightning patterns — structured rewards, triplet learning, prompt optimization, per-agent metrics

Performance TrackingLayer 6: Evaluation

Multi-dimensional scoring — correctness, efficiency, cost, quality per task per agent

Dual-Gate ReviewLayer 7: Security

Gate 1: Security (no telemetry, no keys, no exfiltration). Gate 2: Quality (maintained, documented, useful).

Ruflo Security EngineLayer 7: Security

CVE detection, threat modeling, vulnerability scanning, input validation at system boundaries

PaperclipLayer 8: Company

Company orchestration — org charts, department workflows, budgets, governance for multi-agent teams

Claude PeersLayer 9: Network

Cross-instance messaging — Claude sessions discover and coordinate across machines and projects

Hive MindLayer 9: Network

Ruflo consensus coordination — Byzantine fault tolerance, gossip protocols, distributed decision-making

Why This Structure Exists

This isn't just folder organization. Each component solves a specific problem that makes AI unpredictable without it.

CLAUDE.md is the single source of truth

Without it, every session starts from zero. With it, AI has persistent identity, rules, and context from the first message.

Skills are modular files, not inline prompts

Reusable across projects and sessions. Version-controlled. No more copy-pasting the same instructions.

Hooks are separate from commands

Hooks are deterministic safety checks (always run). Commands are triggered workflows (run when asked). Mixing them creates unpredictable behavior.

Memory has 7 tiers, not 1

Session context expires. Working memory gets pruned. Knowledge persists forever. Different lifespans need different storage — just like human memory.

The auditor gates knowledge promotion

Not every observation deserves permanent storage. The auditor reviews nominations before they become rules — preventing knowledge base bloat and contradictions.

Agent-to-agent communication via files

Agents hand off context through shared memory files, not fragile API calls. File-based state is inspectable, debuggable, and survives crashes.

Key Files & Directories

CLAUDE.md# Your AI brain's operating manual — identity, rules, architecture
memory.md# Active session context — what you're working on right now
knowledge-base.md# Learned rules and patterns that persist across sessions
settings.json# Hooks, permissions, and MCP server configuration
~/.claude/skills/# Directory of 1,730+ skill instruction files
.claude/agents/# 109 agent definitions — OpenClaw specialists + Ruflo swarm agents
.claude/commands/# Slash commands + 10 Ruflo command groups (sparc, github, analysis, etc.)
.claude/hooks/# Safety and quality enforcement scripts
.claude-flow/# Ruflo swarm config, sessions, and runtime data
.mcp.json# MCP server registry — Ruflo (310+ tools), code-review-graph, peers

7-Tier Memory Architecture

See the full deep-dive in the memory system guide.

TierScopeStorageRetention
1. SessionCurrent conversationContext windowUntil session ends
2. WorkingActive taskmemory.mdPruned weekly
3. KnowledgeSystem rulesknowledge-base.mdPermanent (auditor-gated)
4. AgentPer-agentagent-memory/Permanent
5. GraphStructured relationsMCP knowledge graphPermanent
6. VectorSemantic embeddingsRuflo HNSW + SQLitePermanent
7. ArchiveDaily logsmemory/ directoryPermanent

Self-Improvement Loop

This system doesn't just follow instructions — it learns from experience and improves its own operating procedures. Here's how:

1
ObservationAgent encounters a new pattern, error, or useful technique during work
2
NominationLearning is added to knowledge-nominations.md as a candidate rule
3
Auditor ReviewThe Auditor agent reviews the nomination for accuracy and value
4
PromotionApproved rules are promoted to knowledge-base.md — permanent system knowledge
The system also runs /retro retrospectives and tracks performance in PERFORMANCE.md — creating a continuous evaluation and improvement cycle. Every session makes the next one better.

Evaluation & Observability

Production AI systems need monitoring. Our architecture includes built-in tracing, audit trails, and quality gates.

Audit Trail

Every hook logs actions to .claude/logs/. Full history of what the AI did and why.

Quality Gates

The Auditor agent reviews work quality before it ships. The /audit command triggers on-demand reviews.

Performance Tracking

Token usage, task completion rates, and context health monitored across sessions via PERFORMANCE.md.

Get the Full Blueprint — Pre-Configured

Our Pro package delivers this entire system ready to run. One command, 5 minutes, fully operational.

View Pricing