Cornelius
Compounding intelligence your company owns — the cognitive architecture that makes your work compound, built on Claude Code + Obsidian and run on Trinity.
Cornelius turns what you and your company do into knowledge you keep and build on — without retraining a model. It captures your thinking in your own voice, finds the connections you missed, works on open questions on a schedule, and stops every conclusion at one human gate before it becomes part of what you believe. Open source under the MIT license: github.com/Abilityai/cornelius.
Every note records who authored the thinking (you, an external source, or the AI). Nothing the AI concludes becomes endorsed knowledge without your explicit act, and every change leaves a git trail. Adaptive improvement inside strict governance — growth is a dial, with your hand on it.
Cornelius AI Agent — Architecture Explained
Jul 2026
How Cornelius Works
The layer cake, the vault, the capture-to-endorsement loop, Local Brain Search and the Brain Dependency Graph, provenance, reference scopes, and the sub-agents
Getting Started with Cornelius
Run Cornelius standalone in Claude Code with Obsidian in five minutes, add optional MCP servers, then deploy it to your own Trinity for autonomous operation
Using Cornelius
Day-to-day workflows — advise, decide, recall, find connections, capture and interview, research, write — with example prompts and two worked routines
Cornelius Skills
Every one of Cornelius's 60 skills, grouped by job: capture, thinking and deciding, search and connections, autonomous thinking, research, reference data, writing, the graph, and maintenance
What it's for
Cornelius is designed to be the compounding intelligence engine inside Trinity. On Trinity it becomes a persistent brain rather than a chat session, in one of two roles:
The place your organization's experience accumulates and stays: decisions, research, records of people, clients, and competitors, and the reasoning behind them. Other agents in your Trinity fleet consult it, so what one agent records, the others can build on — and it does not walk out the door when a person does.
Your own thinking, captured in your own voice, connected across everything you read, and kept working on your open questions while you are away.
What it does
Project Cornelius= Claude Code + custom agents + Obsidian + FAISS vector search + a governed autonomous thinking loop. It's like having a highly specialized AI research assistant that:
/decide structures real choices with a named decision rule and tripwiresShips pre-seeded
Cornelius isn't an empty template. It comes with a working knowledge graph of ~1,000 interlinked notes on decision-making, judgment, and cognitive science (distilled from published research and books, ~5,000 edges). Clone it and /advise, /recall, or /find-connections work immediately. Start at MOC - Knowledge Base, then layer your own thinking on top. See knowledge-base-analysis.md for what's inside.
Why it is yours
It runs on your machine or your own Trinity server, on plain Markdown files, against any model — export it and walk away. The more a system improves itself, the more it matters who owns it: compounding intelligence on rented ground compounds for the landlord.
Where it comes from.Ability AI is the lab for AI-native companies, and Cornelius is its cognitive architecture for making a company's work compound. It runs in production on Ability AI's own work. The repository is its open template: the memory engine, the governance contracts, and the everyday thinking tools ship there as they stabilize; components of the autonomous loop are released once they have soaked in production.
Why it runs on Trinity
Trinity supplies what a compounding brain needs and a laptop does not: always-on scheduled loops (incubation, domain watch, research), agent-to-agent delegation, an isolated container with your data on your infrastructure, the live Brain Orbvisualization, and receipts — git history, audit logs, and human gates on every change.
Cornelius works standalone in Claude Code for exploration and development. It is meant to be run on Trinity — your own, self-hosted instance. See Deploy to your Trinity.
Latest version
The current version is v10.26: /decide, shared reasoning checks, read roles, calibrated similarity scores, and 60 skills. Every Cornelius version has its own page on the shared changelog — What's New in Cornelius v10.26, and the rest in What's New.