I Fired My Second Brain

About a year ago, Stephen's company built what he still believes was the right AI knowledge system for the tools of the time: a dedicated machine, a folder per client, Notion as the front end, and two AI memory products. Then the tools moved. In this weekend edition, Stephen explains why his company retired its "second brain," what replaced it, and how to run the same audit in your own company.

In this episode, Stephen Forte covers:

Three questions. Why does something exist, is it truly necessary, and is there a simpler or faster way to the result?

The audit. One fact about one client sitting in about eight places, a repair job that creates conflicts, and a dashboard opened for exactly one thing.

Where AI memory products fit. In his team's experience, they went stale and cost more upkeep than they were worth, at least for them. Letta's vendor benchmark and Andrej Karpathy's plain-text approach point the same way.

Three boxes. What you write goes in Git, what you receive is stored and never edited, and what you search is a disposable index. A fast librarian, not the library.

Firewalls and data residency. Each client gets its own container and index, with in-region setups for the UK, the Middle East and the EU when compliance requires it.

Not tied to one AI. Knowledge in neutral places means the best tool for each task, with shared instructions through the AGENTS.md convention.

Your weekend job. A fact census, the three-box sort, the quiet failure test, the reread test, and a one-screen brief for your top customer on Monday.

Sources:

Letta, "Benchmarking AI Agent Memory": https://www.letta.com/blog/benchmarking-ai-agent-memory/

Andrej Karpathy, LLM wiki gist: https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f