◇ Could this help me?
Yes, and the routing idea is the part that is genuinely underrated. An agent with no map searches, and searching is where context and tokens go before any work starts, so a short file that says where to look for what pays for itself immediately and requires no infrastructure at all. The folder-level notes-to-the-next-agent pattern is the same idea applied locally and works for the same reason. General practice, with two honest qualifications: the ninety-five percent saving is one person's measurement of one setup and should not be expected, and the dismissal of retrieval systems is overstated rather than wrong, since structured retrieval does earn its place once the material is larger than a human-maintainable folder tree. Start here, and add retrieval only when the file approach visibly fails.
Help me build a workspace instead of a knowledge system. Based on the kinds of work I describe below, propose a folder structure for you to work in, then write the routing file that sits at the top: short, and saying where to go for each kind of request rather than describing everything. Add a note inside each folder telling the next agent working there what to record so the same thing is not relearned. Keep the routing file as short as it can be while still being unambiguous, and tell me the instruction count. Then tell me honestly at what point this approach would stop working for me and a real retrieval system would be justified, so I know what signal to watch for.