From SDD to MDD: the project now works on top of an engineering memory

Since May, every change to Conn2Flow goes through a written process: human request, small batch, validation and decision record. With so much work done by AI agents in parallel, the heaviest part was no longer the process but the memory: what was decided, what went wrong and why. Hence the new name, Memory-Driven Development.

What changed in practice

• The sdd/ folder became memory/ in the core, the site and the other repositories of the ecosystem, keeping the history (the move was done as a rename in Git).
• Each repository has three base files: reference architecture, general memory and policy. Two engineering memories sit next to them: the Chief Engineer's, which no agent edits, and the execution memory, which agents update at the end of each task.
• Control files keep at most ten active items; the rest goes to archive/, with a one-line summary and the link.

Roles and autonomy

Work is split between the Architect, who writes the request and decides, the Executor, who implements and validates, and the Reviewer, who audits. Each request's setup says how many take part (solo, pair or triad), and the autonomy level says how far an agent goes on its own: supervised, monitored autonomous or background autonomous. At no level does an agent publish to production, and only the human signs off the acceptance.

The triple lock

New artifacts are born with a YAML header (identifier, state, summary). Each folder's index is generated from those headers, and handing a batch between roles carries the project, request, batch, scope, evidence and open items. The shared catalog has 45 skills, synced from a single source.

Status

The migration is on main in the core and the site, with the core's full suite green (1,768 PHP tests and 664 JavaScript tests). The overall acceptance by the Chief Engineer is pending.