The scalability problem with manual PKM

At some point in my use of AI agents, I found it easier to move from the desktop app to GitHub and the terminal, because that way I could organise conversations into different folders as text files on my local drive, keep them as .md notes, and review them in Obsidian whenever I needed. Since then I've kept a git repo as my knowledge base, managed by an AI agent running in the terminal attached to that folder, automatically writing notes into subfolders organised by conversation or project.

After generating around 1,500 AI conversation notes that way, I came across OptMem and applied it on top. It runs a series of automatic passes where an AI manager organises memories into layers, so the system can access a large corpus without flooding the context window. I've also looked into IWE, which others in this thread have mentioned — it looks promising, though as far as I can tell it's less automated than OptMem and relies more on the human doing the structural work upfront.

Which brings me to something I keep thinking about: how scalable non-automated PKM really is. How much time can we realistically keep putting into manually maintaining a collection that only grows — compared to what else we could be doing with that time, like actually using the PKM to produce original writing or research?

Cheers, and may your backlinks resolve,
Zhenia