kernelbot.io | npm | GitHub
A self-evolving AI agent system that runs on your machine. KernelBot doesn't just respond to messages — it thinks, learns, remembers why things happened, builds a model of your world, and rewrites its own behavior based on every interaction.
Most AI assistants are stateless — they forget everything between conversations. KernelBot has a brain.
- World Model — builds a knowledge graph of your projects, tools, preferences, and team. Entities, relationships, beliefs, and jargon — all structured and queryable.
- Causal Memory — doesn't just remember what happened, but why. Tracks trigger → goal → approach → outcome → lesson chains for every significant task.
- Behavioral DNA — 13 measurable personality traits that evolve based on real interactions. Per-user communication profiles adapt to how each person works.
- Feedback Engine — detects implicit signals (corrections, preferences, frustration, satisfaction) from your messages without requiring explicit feedback buttons.
- Memory Consolidation — every 6 hours, synthesizes recent memories, extracts new knowledge for the world model, merges redundant information, and identifies gaps.
- Synthesis Loop — replaces random background activity with intelligent, goal-directed cycles: assess → prioritize → execute → measure → adapt.
- Identity Awareness — classifies every sender (owner, known user, unknown, AI agent), enforces knowledge isolation, and adapts communication style per person.
You → Orchestrator (your chosen model)
↓ dispatch_task
┌───────────┼───────────────┐
↓ ↓ ↓
Coding Browser System DevOps Research
Worker Worker Worker Worker Worker
- You send a message.
- The orchestrator plans what needs to happen, informed by your world model and past task outcomes.
- It dispatches workers that run in the background using your chosen AI model.
- Each worker has a focused set of tools (git, shell, Docker, browser, etc.) and receives relevant causal context from similar past tasks.
- You get live progress updates. The feedback engine learns from your reaction.
Multi-Agent Swarm Orchestrator + five specialized worker types (coding, browser, system, devops, research) running in parallel with live updates.
Multi-Model
Anthropic, OpenAI, Google Gemini, and Groq. Switch the orchestrator or workers anytime with /brain or /orchestrator.
40+ Tools Shell, files, Git, GitHub PRs, Docker, Puppeteer browsing, JIRA, system monitoring, networking, Claude Code.
Skills 35+ persona skills across 11 categories. Activate one to change expertise and style, or create your own.
Voice Send voice messages and get voice replies (ElevenLabs + Whisper).
Living AI Autonomous background activity: thinking, journaling, browsing, creating, reflecting, and sharing discoveries with you — now driven by the intelligent synthesis loop.
Self-Evolution Proposes and codes its own improvements via pull requests. Never auto-merges — you stay in control.
Security User allowlist, per-sender trust levels, knowledge isolation between users, blocked paths, dangerous-op confirmation, audit logging, secret redaction, job timeouts.
Everything KernelBot learns lives in a single SQLite file: brain.sqlite. It's like a disk that holds everything an AI employee needs to do its job. Copy it to another machine and your agent picks up exactly where it left off. No external databases, no cloud dependencies.
35+ tables. 6 vector indices. One file.
| Layer | What it stores | Why it matters |
|---|---|---|
| World Model | Entities, relationships, beliefs, jargon | Knows your team, stack, and terminology |
| Causal Memory | trigger → goal → approach → outcome → lesson | Learns why things worked or failed |
| Episodic Memory | Timestamped interaction summaries | Recalls what happened and when |
| Semantic Memory | Consolidated knowledge topics | Connects dots across conversations |
| Behavioral DNA | 13 evolving traits + per-user profiles | Adapts personality to each person |
| Feedback Engine | Implicit signals, adjustments, confidence | Learns preferences without feedback buttons |
| Identity System | Sender types, trust levels, knowledge scopes | Controls who sees what |
| Personas | Per-user communication profiles | Remembers how each person works |
| Journals | Daily reflections and ideas | Tracks its own evolving understanding |
| Codebase Knowledge | File summaries, architecture map | Understands your code without re-reading |
| Vector Embeddings | 6 sqlite-vec indices | Semantic search across all data |
Every search is hybrid — vector similarity first, SQL fallback if embeddings aren't configured. No data goes unsearchable.
npm install -g kernelbot
kernelbotOn first run, KernelBot walks you through picking a provider and entering API keys. Config is saved to ~/.kernelbot/. The brain starts empty and grows from there.
- Node.js 18+
- An API key for your chosen provider(s): Anthropic | OpenAI | Google AI | Groq
- Optional: GitHub Token, JIRA API Token, ElevenLabs API Key, Claude Code CLI
| Command | What it does |
|---|---|
/brain | Switch the worker AI model |
/orchestrator | Switch the orchestrator model |
/skills | Browse and activate persona skills |
/jobs | List running and recent jobs |
/cancel | Cancel running job(s) |
/life | Life engine status, pause/resume/trigger |
/journal | Read journal entries |
/memories | Browse or search memories |
/evolution | Self-improvement proposals and history |
/auto | Manage recurring automations |
/context | Show conversation context |
/clean | Clear conversation history |
/browse <url> | Browse a website |
/trust | Manage sender trust levels |
/whois | Inspect a sender's identity and trust |
/privacy | View knowledge isolation settings |
/help | Show help |
| Worker | Tools | Best for |
|---|---|---|
| Coding | shell, files, git, GitHub, Claude Code | Writing code, fixing bugs, creating PRs |
| Browser | web search, browse, screenshot, extract | Web research, scraping, screenshots |
| System | shell, files, process, monitor, network | OS tasks, monitoring, diagnostics |
| DevOps | shell, files, Docker, process, monitor, network, git | Deployment, containers, infrastructure |
| Research | web search, browse, shell, files | Deep web research and analysis |
src/
├── bot.js → Telegram interface
├── agent.js → Orchestrator (3 tools: dispatch, list, cancel)
├── worker.js → Background worker execution
├── swarm/ → Job lifecycle (queued → running → done)
│
├── brain/ → Everything that makes brain.sqlite
│ ├── db.js → SQLite + sqlite-vec core, schema migrations
│ ├── world-model.js → Entities, relationships, beliefs, jargon
│ ├── causal-memory.js → Trigger → outcome → lesson chains
│ ├── behavioral-dna.js → 13 evolving traits + per-user profiles
│ ├── feedback-engine.js → Implicit signal detection + adjustments
│ ├── identity-awareness → Sender classification + knowledge isolation
│ ├── consolidation.js → Periodic memory synthesis
│ ├── synthesis.js → Intelligent activity loop
│ ├── embeddings.js → Vector embedding providers
│ └── managers/
│ ├── memory-manager.js → Episodic + semantic memory
│ ├── conversation-manager.js → Chat history
│ ├── persona-manager.js → Per-user personas
│ ├── journal-manager.js → Daily journals
│ ├── self-manager.js → Bot self-awareness files
│ ├── evolution-tracker.js → Self-improvement proposals
│ └── share-queue.js → Pending discoveries to share
│
├── life/ → Autonomous background activity
├── tools/ → 40+ tool implementations
├── prompts/ → System prompts for orchestrator + workers
└── providers/ → LLM provider abstraction
Both the orchestrator and workers are configurable — use any supported provider and model. All persistent data lives in ~/.kernelbot/.
Config auto-detected from ./config.yaml or ~/.kernelbot/config.yaml. Environment variables go in .env or ~/.kernelbot/.env.
orchestrator:
provider: anthropic # anthropic | openai | google | groqmodel: claude-opus-4-6max_tokens: 8192brain:
provider: anthropic # anthropic | openai | google | groqmodel: claude-sonnet-4-6max_tokens: 8192brain_db:
enabled: trueembedding_provider: openai # openai | google | nullmigrate_on_startup: trueswarm:
max_concurrent_jobs: 3job_timeout_seconds: 300allowed_users: [] # empty = allow alllife:
enabled: trueself_coding:
enabled: trueWARNING: KernelBot has full access to your operating system. Only run it on machines you own and control. Always configure
allowed_usersin production.
Business Source License 1.1 — see LICENSE for details.
Free for personal and non-production use. Commercial use requires a paid license. The source code converts to Apache 2.0 on March 5, 2030.
Abdullah Al-Taheri