Glean's "context graph" framework for corporate information handling has really got me thinking about how to level up my personal knowledge management game.
I'm not 100% sure about the way to approach it either, and I've been figuring out the connection between enterprise software use of log collection in services like Grafana and how that might related to the metadata in our PKM. As you said in your piece, breadcrumbs could be key here. If we added structured data to this (like many log collections do), then AI could help in constructing useful historical trails. Or possible pathways for us to review, much like an actual trail and the decisions we make in which path to take.
Maggie never disappoints! Now that you’ve reminded me I remember reading this a year or so ago, but my brain was all foggy from the baby and I didn’t take good enough notes :P
Structured data is really important but so hard to have the discipline to do, even with AI to help.
I’m actively avoiding using ‘AI’ tools, but still found this post interesting.
Apart from in metadata, search history, etc, much of my ‘how did I get here?’ data will be recorded in an unstructured way in my (electronic) work Journal, where I record my current thinking, what I plan to do about it, and the immediate outcomes. Other people’s might be found in their ‘Daily Notes’.
I never read through my past Journal entries, as I use them for ‘thinking out loud’ to myself to overcome current issues, but perhaps I should be reviewing them occasionally. (That said, I can’t imagine anything more off-putting than reading a historical litany of seldom-met good intentions!)
I generally do go back through my notes at least once or twice, if only to hunt down old ideas and tasks when I have some free time, but I don't usually "think out loud" in my notes, so it's sort of the opposite. I probably should do more "working through things" on paper, but I usually think it through in my head and then write down the plan... or actively discuss with someone out loud or in conversation. Not sure how to change that habit, really, and I don't like the idea of feeding even typed conversations into an AI to summarize.
Your post really got me thinking over the last 24 hours. I was
surprised how much Claude helped me think about this. I initially got an idea of
adding new properties that might help discovering the process gaps. Since the Claude/API lacks access to revision history, Claude suggested adding a 'trajectory' property (e.g., converging, open, or stalled). While subjective, externalizing that implicit knowledge allows for better data layering
and provides a clearer birds-eye view of my workflow. And it could assist Claude identify, in combination with other properties, the "good stuff" you allude to. Definitely food for more
An easy solution to Claude lacking access to revision history might be to use git -- I've gotten a lot of mileage out of frequent commits even though I don't understand git hardly at all... because Claude does :D
I have a couple of plugins for it that help sort of “omni-search” the database, but mostly it’s just a lot of years put into it and always being meticulous about cross-linking things together before any problems start haha :D
I found this hugely interesting and of course I can’t wait to integrate something that resembles your process graph idea into my notes/productivity app with MCP that I’m building.
This project is getting cooler by the minute, and inspiring posts like this one are absolutely vital to help me set the direction🙏
Not yet, I’m afraid… I took longer than I hoped getting it ready for production and I had this down for phase 2. And of course some other shiny thing caught my fancy first (remarkable note sync!)…
That said I’ll look into now. Feel free to send any suggestions my way 😎.
The main tip is that Claude is a fast learner and if you tell it what you want it can figure it out pretty well if you just keep hammering away at emphasizing what you want, don't let it talk you out of it.
over the last year or so i have done some extensive work with claude for my ttrpg gamemaster and player notes. most recently i had it create a skill that reads my last player session note and from that creates new atomic notes either from deadlinks i created or from inferred meaning. It has a copy of my live vault and compares existing notes and determines if there is a need for updating or creating something new. I gave it no real instructions about what to do, but it puts in a header of the session number and writes a small recap from that session in the respective note (npc, location, item, lore, faction) and it puts in links to other notes that mentions the note and updates meta data. It even created new meta data for me that fit the vaults structure. As a player that saves me time, i dont have to go through my session notes creating new notes from all my deadlinks and fill out templates etc.. So i have actually found something where AI gives me less work to do :) but i think its akin to innovation. You need an idea and you need to iterate through it to get anywhere. Much like other automatisation, it takes alot of work to set it up just right, but when you get out on the other side of that, thats where it begins helping. The "setup" is iterating through ideas and instructions to see what works and how until you distill the correct process for that specific idea.
Iteration is _absolutely_ a key part of making all this stuff actually work and be useful. One of the nice things about LLMs is that they make the iteration loop tighter, actually.
One idea I keep coming back to when thinking about the specific pathways we use to reach insight is historical trails.
https://maggieappleton.com/historical-trails
I'm not 100% sure about the way to approach it either, and I've been figuring out the connection between enterprise software use of log collection in services like Grafana and how that might related to the metadata in our PKM. As you said in your piece, breadcrumbs could be key here. If we added structured data to this (like many log collections do), then AI could help in constructing useful historical trails. Or possible pathways for us to review, much like an actual trail and the decisions we make in which path to take.
Maggie never disappoints! Now that you’ve reminded me I remember reading this a year or so ago, but my brain was all foggy from the baby and I didn’t take good enough notes :P
Structured data is really important but so hard to have the discipline to do, even with AI to help.
I’m actively avoiding using ‘AI’ tools, but still found this post interesting.
Apart from in metadata, search history, etc, much of my ‘how did I get here?’ data will be recorded in an unstructured way in my (electronic) work Journal, where I record my current thinking, what I plan to do about it, and the immediate outcomes. Other people’s might be found in their ‘Daily Notes’.
I never read through my past Journal entries, as I use them for ‘thinking out loud’ to myself to overcome current issues, but perhaps I should be reviewing them occasionally. (That said, I can’t imagine anything more off-putting than reading a historical litany of seldom-met good intentions!)
I generally do go back through my notes at least once or twice, if only to hunt down old ideas and tasks when I have some free time, but I don't usually "think out loud" in my notes, so it's sort of the opposite. I probably should do more "working through things" on paper, but I usually think it through in my head and then write down the plan... or actively discuss with someone out loud or in conversation. Not sure how to change that habit, really, and I don't like the idea of feeding even typed conversations into an AI to summarize.
Your post really got me thinking over the last 24 hours. I was
surprised how much Claude helped me think about this. I initially got an idea of
adding new properties that might help discovering the process gaps. Since the Claude/API lacks access to revision history, Claude suggested adding a 'trajectory' property (e.g., converging, open, or stalled). While subjective, externalizing that implicit knowledge allows for better data layering
and provides a clearer birds-eye view of my workflow. And it could assist Claude identify, in combination with other properties, the "good stuff" you allude to. Definitely food for more
thought.
Thanks!
An easy solution to Claude lacking access to revision history might be to use git -- I've gotten a lot of mileage out of frequent commits even though I don't understand git hardly at all... because Claude does :D
I have just shy of 4 million words in my Obsidian database, almost all written by me.
The amount of things I have cross-linked, patched together, and nailed in place is absolutely astounding after this almost 6 years of work.
Automation in this kind of context? Not going to happen hah... there's just too much.
But I can say, every time I go digging, I 'glean' something new I'd never really put together before, and that's always exciting.
Do you ever consider layering on better search tools, even if you're not interested in automations?
I have a couple of plugins for it that help sort of “omni-search” the database, but mostly it’s just a lot of years put into it and always being meticulous about cross-linking things together before any problems start haha :D
I found this hugely interesting and of course I can’t wait to integrate something that resembles your process graph idea into my notes/productivity app with MCP that I’m building.
This project is getting cooler by the minute, and inspiring posts like this one are absolutely vital to help me set the direction🙏
Did you ever end up integrating any of this? 👀
Not yet, I’m afraid… I took longer than I hoped getting it ready for production and I had this down for phase 2. And of course some other shiny thing caught my fancy first (remarkable note sync!)…
That said I’ll look into now. Feel free to send any suggestions my way 😎.
The main tip is that Claude is a fast learner and if you tell it what you want it can figure it out pretty well if you just keep hammering away at emphasizing what you want, don't let it talk you out of it.
over the last year or so i have done some extensive work with claude for my ttrpg gamemaster and player notes. most recently i had it create a skill that reads my last player session note and from that creates new atomic notes either from deadlinks i created or from inferred meaning. It has a copy of my live vault and compares existing notes and determines if there is a need for updating or creating something new. I gave it no real instructions about what to do, but it puts in a header of the session number and writes a small recap from that session in the respective note (npc, location, item, lore, faction) and it puts in links to other notes that mentions the note and updates meta data. It even created new meta data for me that fit the vaults structure. As a player that saves me time, i dont have to go through my session notes creating new notes from all my deadlinks and fill out templates etc.. So i have actually found something where AI gives me less work to do :) but i think its akin to innovation. You need an idea and you need to iterate through it to get anywhere. Much like other automatisation, it takes alot of work to set it up just right, but when you get out on the other side of that, thats where it begins helping. The "setup" is iterating through ideas and instructions to see what works and how until you distill the correct process for that specific idea.
Iteration is _absolutely_ a key part of making all this stuff actually work and be useful. One of the nice things about LLMs is that they make the iteration loop tighter, actually.