Session capture turns ephemeral AI conversations into a compounding knowledge base

shadcn’s innovation is a /done skill that runs after every Claude Code session, capturing key decisions, questions asked, follow-ups needed, and context — dumped to a markdown file tagged with session ID and branch name. “Helpful when I need context later” understates the effect: it creates a searchable development memory that compounds over months.

This is file-based memory architecture (the Resources → Items → Categories pattern from Persistent agent memory preserves institutional knowledge that walks out the door with employees) applied to the development workflow itself. Raw conversations are Resources, extracted decisions and follow-ups are Items, and project-level documentation is the Category layer. It extends Spec files are external memory that survives context resets beyond project specs to capture the reasoning and context that specs omit — the “why did we consider and reject X” that Decision traces are the missing data layer — a trillion-dollar gap identifies as a trillion-dollar gap. The pattern is lightweight enough to sustain daily, which is what makes Compound engineering makes each unit of work improve all future work practical rather than aspirational. In Chase’s taxonomy of Hot-path and offline learning are two temporal modes for agent context updates — each with different tradeoffs, /done is hot-path + explicit learning — the user triggers it during the session, and learnings are captured immediately rather than extracted from traces after the fact.