A conference talk by Magnus Gille presenting a framework for making AI-assisted development compound over time rather than just speed up individual tasks. The framework rests on three pillars: persistent context memory (a self-hosted Raspberry Pi system called Munin using MCP, with a status.md file per repo), adversarial review (having Claude Code invoke Codex in a headless mode to debate and critique outputs across multiple rounds), and defining 'good' by instrumenting metrics to make quality measurable and improvable. The talk covers practical implementation details including how sessions are opened and closed, how memory is pruned via time-decay ranking and nightly synthesis, and how token costs are managed by scoping sub-agent prompts tightly.