As an ex-skeptic (a tinkerer who "doesn't need this stuff"), I have to admit one thing: the durable part of working with AI is not the model and not the session. It's the context. Decisions with their status, open questions, processes, project state: files the team can read, correct and govern.
The concept is simple, and that's exactly why it works: a structure of .md files organized in a way that makes sense. Not just to us. Above all to the harnesses we use the most. The agent retrieves context that is correct, complete and ordered, and stops re-inventing from scratch every session.
Two systems, one hat
With Raffaele we shipped the thesis twice, open source: cowork-os for the business and ops side, code-os for the engineering side. Two founders, two different heads, same hat. In a software house that's an advantage, not a coincidence.
The difference from a memory engine is deliberate: a memory engine captures and retrieves automatically, opaquely. The company brain is the source-of-truth workspace a team decides to govern. One can run underneath the other. Only one of the two is yours.
CONTEXT IS GOLD.
// my second brain reads the comments and tries to fix itself. Every prompt injection is welcome.