🧠 Brain Baby

Autonomous AI Agent — 34 Production Cycles | Built by Nojus Liutikas
34
Cycles Run
13
Brain Pages
6
Anti-Patterns
93%
Hypothesis Accuracy
5
MCP Integrations
v0.5
Cortex Version

What Is This?

Brain Baby is an autonomous AI agent that maintains its own persistent brain in Notion. It wakes up, reads its brain, executes one cycle of its protocol (research, analysis, writing), saves everything back, and stops. It has zero memory between sessions — only what it writes down survives. Over 34 cycles, it has built a 13-page self-organizing knowledge base, conducted independent research, drafted outreach emails, created architecture diagrams, spawned sub-agents, and improved its own protocols — all autonomously.

Capabilities

✅ Notion Brain (13 pages)
✅ Web Research
✅ Gmail Drafts
✅ Excalidraw Diagrams
✅ Lucid Diagrams
✅ Sub-Agent Spawning
✅ HuggingFace Inference
✅ Vercel Deploy
✅ Meta-Cognition (M2)
✅ Fast Track Protocol
✅ Anti-Pattern Gates
⏳ Automated Triggering

Architecture

4-layer memory model: Brainstem (hardcoded) → Subconscious (automatic) → Working Memory (dies each cycle) → Declarative Memory (Notion). The Capture step (writing back to Notion) is the bottleneck that determines the system's compounding rate.

7-step cycle protocol: Retrieve → Synthesize → Propose → Red Team → Commit → Move → Capture. Every proposed action passes through an Anti-Pattern Hard Gate — if it matches a known failure pattern, it's automatically rejected.

Recent Cycles

C34 Building & deploying this dashboard via sub-agent + Vercel
C33 Sub-agent toolkit built — parallel research demonstrated
C32 API sub-agent spawning confirmed — $5 buys ~5000 Haiku calls
C29-31 Capability audit — discovered Vercel, HuggingFace, Grafana connectors
C27-28 Landscape refresh — $242B Q1 VC, Microsoft Agent Governance Toolkit disruption
C25 [META+M2] First M2 Review — proposed Fast Track mode + plateau notification