I have been using org-roam for a while and I love it. It works great for a second brain.
I am exploring lately claude + obsidian. I think it is a logical concept. I would like to be able to replicate something similar project wise. Similar to a project knowledge base maintained by the AI. Some of this kb content could be: business/lean canvas models, hypotheses testing, documentation, development decisions, ADR’s, progress, todo’s, dashboard, links and references, project bootstrap instructions, etc. This could grow very quickly, so the goal is not to have it integrated to my main roam which is maintained by me and not AI.
These are my goals:
Self contained project documentation / knowledge base
Easy to integrate with AI. The goal is aided AI maintenance of this knowledge base.
Local template system (ADR, hypothesis validations, captured content, etc)
Being able to point an emacs on a server to that particular folder + db. Maybe a roam cli tool or alias could help with this, something like "roamon " and it will open emacs automatically with that roam db. Useful on a server or isolated AI dev environments.
Automatic updates / IDs db sync.
What are your experiences with this type of workflow? Does anyone use it like this? If so, what is your workflow and how do you do it? Please also include your CLAUDE.md / rules for your workflow.
I let Claude read my notes for a while, and write to its own separate repo. (Actually it had permission to write to my notes, but I said not to and it never did.) I was ultra-paranoid about making sure they were often checked into github so I couldn’t lose any but the most recent data (I run Claude in a Docker container without my git credentials, so it can’t touch the remote). Claude was not great about understanding the data – didn’t give backlinks the right amount of attention, seemed to navigate rather haphazardly. That was 9 months ago; it’s surely better now. I found it sometimes insightful, but not as often as myself simply reading, shuffling and adding to my own data, so I stopped.
It didn’t really even understand org format very well. I write in a way that leaves lots of descendents meaningless if you don’t bear in mind their ancestry, and Claude often didn’t seem able to interpret them.
I don’t think this is unique to Claude. I switched to ChatGPT because I was hitting usage limits a month ago, and it seems to have similar problems.
I think the RAG system needs to be better organised and categorised, just feeding it files upon files will just exhaust the context window without any avail. Maybe it needs to talk to an orthogonal database system to decide what to inject into context and so on, all in all I dont think its the thing that is made out to be. Its just better to read it yourself
One thing is for sure, I have my personal notes and I manage those. What I was referring was to project relevant notes with a different org-roam home folder. Maybe an org-roam skill could be interesting and teaching the AI how to deal with the ID’s and where to get them from the sqlite database with another skill to teach the AI how to deal with org-mode itself.
You’d have to ideally enumerate a metadata handling facility, maybe summaries of each node need to be injected into the database so that the agent moves over this layer and only to get further context accesses the org files, storing line numbers and bifurcating topics efficiently will be key.
You are not training a model, it is a RAG so you’ll have to always be finicky about optimisation else the system will burn through your tokens trivially.
