[FD#03] The Step You're Not Designing

Most PKM advice is capture advice. Which app to use, how to tag things, PARA vs Zettelkasten. The retrieval step barely comes up.

That asymmetry made sense before AI. You had to design capture well because retrieval depended on it. Bad structure in, bad retrieval out. But the dynamics are changing.

Between capture and retrieval, there is now a third thing: an automated layer that processes, compresses, and creates linkages before retrieval happens. Tools like OptMem (VictorTaelin, GitHub) compress AI session notes into fixed-width records with binary tree compression. The design claim is "position is identity": the record structure carries meaning without needing a separate index.

What that creates is a different architecture. Not capture then retrieval, but capture then processing layer then retrieval. If the processing layer is automated, retrieval can be designed independently: one user-facing layer (how you ask questions) and one database-facing layer (how the system navigates the memory).

That is the design problem worth spending time on. Not "how do I capture better?" but "how do I build a retrieval mechanism that works at scale?"