I’ve been building a small tool called oamc around a workflow I wanted for personal research and long-running project memory.
The basic idea is: instead of repeatedly querying raw notes/documents, sources get ingested into a maintained markdown wiki. The wiki becomes the working knowledge layer, and future questions are asked against that layer instead of against raw text every time.
The pipeline is:
- drop or clip sources into an inbox
- ingest them into source, concept, entity, and synthesis pages
- ask questions against the wiki
- save useful answers back as new synthesis pages
A few things I cared about:
- local-first workflow
- markdown as the actual knowledge layer
- inspectable files instead of hidden memory
- lighter than standing up a full RAG stack
- works well with Obsidian, but doesn’t depend on it conceptually
There’s also a small local dashboard and a macOS menubar app so it can keep running in the background.
This was inspired by Andrej Karpathy’s “LLM Wiki” idea. I was basically trying to turn that pattern into something I’d genuinely use day to day.
Repo:
https://github.com/michiosw/oamc
I’d especially love feedback from people here on:
- wiki-first vs RAG-first for personal knowledge
- where this approach starts breaking down at scale
- whether markdown artifacts are actually a better interface for long-term LLM memory than embeddings + retrieval alone
[link] [comments]