Consilium

Consilium is an agent-operated investment fund that researches public companies, debates ideas, makes paper-trading decisions, and learns from what happens afterward.

It is not an AI stock picker or trading dashboard. The experiment is the process around the decision: analysts, a challenger, a portfolio manager, risk controls, post-mortems, and an institutional memory that compounds over time. It never trades real money.

Sector analysts research companies and write investment memos. A challenger looks for weak assumptions and missing risks. A portfolio manager reads both sides and makes the final paper decision. Position monitors revisit open ideas, and a librarian turns completed trades into reusable lessons.

They do not work through an open-ended group chat. Every handoff is a structured artifact that can be reviewed, traced, and challenged later.

The most important part of the project is not the agents. It is what the system remembers.

Consilium keeps doctrine, research, theses, decisions, execution records, post-mortems, and lessons in a versioned institutional memory. Agents can add to it, but important changes to the fund’s rules still require human approval. The goal is to improve the process instead of starting each new session with a larger prompt.

The agents do not control the boundaries. Position sizing, concentration, cash levels, cooldown periods, the tradable universe, and the paper-only kill switch are enforced in code. Portfolio activity is recorded as immutable events so the current state can be reconstructed and audited.

Numbers used in a decision must come from structured market or filing data. Web research can add context, but it is not trusted as the source of financial figures.

This is still in progress. The current build includes:

• Multi-sector research agents
• Memo and counter-memo workflow
• Portfolio-manager decisions
• Paper execution and accounting
• Hard risk and sizing constraints
• Position monitoring
• Post-mortems and reusable lessons
• Versioned institutional memory
• Source trust and citation tracking
• Agent cost and failure monitoring
• Read-only operator terminal
• Agent and execution kill switches

I’m replacing the remaining demonstration data in the terminal, tightening the live read models, and running the system long enough to learn whether the memory actually improves its decisions.