diff --git a/agents/WingedGuardian__genesis-agi/README.md b/agents/WingedGuardian__genesis-agi/README.md
new file mode 100644
index 0000000..6b3f992
--- /dev/null
+++ b/agents/WingedGuardian__genesis-agi/README.md
@@ -0,0 +1,63 @@
+# Genesis AGI
+
+**Genesis** is an open-source autonomous AI agent cognitive architecture built on [Claude Code](https://docs.anthropic.com/en/docs/claude-code) as its reasoning engine. It is not a chatbot, an API wrapper, or a prompt chain — it is a system designed to get fundamentally better the longer it runs alongside a user.
+
+## What it does
+
+Genesis provides four interconnected capabilities that compound over time:
+
+### 🧠 Persistent 4-Layer Memory
+- **Essential knowledge** — long-term facts, preferences, and context about the user
+- **Proactive recall** — surfaces relevant knowledge before you ask
+- **Semantic search** — Qdrant-powered vector retrieval across everything Genesis has learned
+- **Knowledge pipeline** — transforms raw interactions into structured, reusable procedures
+
+### 📚 Closed-Loop Self-Learning
+Outcome classification, causal attribution, and procedure extraction with Laplace-smoothed confidence scores. Genesis tracks its own accuracy and can show the user the receipts.
+
+### 🌙 Background Cognition
+Genesis acts between sessions — researching, reflecting, auditing, and communicating via Telegram while the user is away. Not waiting for a prompt.
+
+### 🔑 Earned Autonomy
+Trust is granted per action category through demonstrated competence. First failure triggers demotion. The user always controls the levers; Genesis earns its scope of action over time.
+
+## Key capabilities
+
+| Skill | Description |
+|---|---|
+| Persistent memory | 4-layer hybrid memory compounding across months |
+| Self-learning | Outcome-driven learning with confidence calibration |
+| Background cognition | Autonomous research and reflection between sessions |
+| Earned autonomy | Per-category trust model, evidence-based |
+| Web research | MCP-powered search and fetch with anti-bot handling |
+| Telegram integration | Async bidirectional channel |
+| Code intelligence | GitNexus blast-radius analysis + Serena LSP |
+
+## Getting started
+
+```bash
+git clone https://github.com/WingedGuardian/GENesis-AGI.git ~/genesis-setup
+cd ~/genesis-setup
+./scripts/host-setup.sh
+```
+
+After install: `cd ~/genesis && claude`
+
+## Requirements
+
+- Ubuntu 22.04+ (dedicated Linux machine or VM)
+- Claude account (Claude Code as reasoning engine)
+- 8 GB RAM minimum, 16 GB recommended
+- Tailscale (for remote dashboard access)
+
+## Runtime
+
+Genesis runs as a set of systemd user units inside an Incus container, with Qdrant for vector storage and a Guardian/Sentinel dual-watchdog for reliability.
+
+## What makes it different
+
+Most AI agents are reactive and stateless — equally ignorant about you on day 100 as day 1. Genesis is different: it accumulates context, acts autonomously, and calibrates its own judgment. It's a cognitive partner, not a tool.
+
+---
+
+[Repository →](https://github.com/WingedGuardian/GENesis-AGI) · [Discord →](https://discord.com/invite/Zkc3XMQpJX) · License: MIT
diff --git a/agents/WingedGuardian__genesis-agi/metadata.json b/agents/WingedGuardian__genesis-agi/metadata.json
new file mode 100644
index 0000000..1c7e812
--- /dev/null
+++ b/agents/WingedGuardian__genesis-agi/metadata.json
@@ -0,0 +1,13 @@
+{
+  "name": "genesis-agi",
+  "author": "WingedGuardian",
+  "description": "Autonomous AI agent with persistent 4-layer memory, closed-loop self-learning, background cognition, and an earned-autonomy framework. Powered by Claude Code.",
+  "repository": "https://github.com/WingedGuardian/GENesis-AGI",
+  "version": "3.0.0-beta",
+  "category": "productivity",
+  "tags": ["autonomous-agent", "persistent-memory", "self-learning", "claude-code", "cognitive-architecture", "background-cognition", "earned-autonomy", "agi"],
+  "license": "MIT",
+  "model": "claude-sonnet-4-5-20250929",
+  "adapters": ["claude-code", "system-prompt"],
+  "icon": false
+}