"I'm Isagi. I devour the field in front of me."
Isagi is a personal autonomous AI operating system — a self-improving, memory-tiered, governance-shaped layer on top of Hermes Agent by Nous Research.
Named after Yoichi Isagi from Blue Lock — the egoist who adapts by understanding the whole field, not just his position. This agent doesn't just execute tasks. It learns, governs itself, and compounds over time.
Inputs → Memory Classifier → L1 Active / L2 Archive → Agent Execution → Governance → Feedback → Future Behavior Change
| Layer | What it does |
|---|---|
| Memory Hierarchy | L1 hot memory + L2 Mem0 archive. Facts scored on write by importance: (usage × 0.4) + (recency × 0.2) + (stability × 0.4) |
| Governance Layer | Behavioral profiles (Builder, Content, Default), failure pattern detection, confidence-gated skill edits |
| Mission Control | 8-directory system: state, knowledge, projects, content, reviews, automation, inbox, governance |
| Confidence Scoring | Every task outcome scored — low confidence blocks blind skill updates |
| GitHub Effect Layer | 5 actions wired into Telegram (issues, PRs, project sync, ship events) |
| Action Engine | 14+ scripts under ~/.hermes/scripts/actions/ — invoke via Telegram /action |
| Self-Improvement | Post-task review → skill creation/update → curator lifecycle management |
| Google Workspace | OAuth2-connected: Gmail, Calendar, Drive, Sheets, Docs, YouTube |
Most people treat AI agents as tools to use. Isagi treats them as kernels to extend.
The difference shows up in the details:
- Memory doesn't fill up because low-importance facts never enter L1
- Skills don't drift because every edit passes a confidence threshold
- The agent doesn't forget core context because stability has 2x the weight of recency
- Nothing is invisible — every session is logged in SQLite with full FTS5 search
- A Linux machine (or VPS — I use Canbato at $10-15/month)
- curl
- A Telegram account
curl -fsSL https://raw.githubusercontent.com/NousResearch/hermes-agent/main/scripts/install.sh | bashThen walk through the setup wizard — model selection, messaging platform (use Telegram), and provider config.
Start by telling Hermes about yourself:
"I am [your name]. I'm a [your role]. Here's what I build, what I'm working on, and my goals."
From that single message, Isagi begins learning.
Download the complete walkthrough (9 chapters, ~12K words):
📄 Isagi: The Egoist AI Agent → PDF
Chapters:
- The Foundation — why Hermes, installation, first message
- The Identity Layer — why "Isagi", the Blue Lock philosophy
- Mission Control — 8 directories, health dashboard, action engine
- The Memory Architecture — L1 vs L2, classifier formula, retrieval pipeline
- The Governance Layer — profiles, failure patterns, confidence scoring
- Self-Improvement — skill engine, review cycle, curator bot
- Practical Use Cases — daily tutor, computer admin, session recall
- Google Workspace Integration — OAuth, Gmail, Calendar, Drive, Sheets, Docs, YouTube
- What I Learned — architecture insights, roadmap
| Component | Tech |
|---|---|
| Agent Framework | Hermes Agent |
| Model | DeepSeek via OpenRouter |
| Memory | Built-in L1 + Mem0 L2 Archive |
| Messaging | Telegram |
| Storage | SQLite (sessions), Markdown (skills/memory) |
| Infrastructure | Canbato VPS |
| Integrations | Google Workspace (OAuth2), GitHub, Tailscale |
- Memory promotion execution — wiring the classifier into the write path
- Kanban execution engine — inbox → classify → decompose → assign → feedback
- Spaced repetition tutor — adaptive difficulty scheduling
- Cross-device admin — deeper Tailscale integration
- Content Engine — auto-repurposing YouTube/trends → LinkedIn/X/Reddit
Built by Rajesh Kalidandi
- LinkedIn: linkedin.com/in/rajesh-kalidandi
- GitHub: github.com/RajeshKalidandi
- Website: rajeshkalidandi.online
Isagi is never finished. That's the point.