building ai knowledge systems with @arsumbrisai
- Tweets 1,294
- Following 357
- Followers 13,466
- Likes 3,840
Pinned Tweet
excited to share a walkthrough of @arsumbrisai
build knowledge bases in typed markdown with your agents via mcp, then build your own apps and plugins on top of it
everything lives in repos you can share and extend. a bit like packages for knowledge work
4
2
30
1,334
excited to share a walkthrough of @arsumbrisai
build knowledge bases in typed markdown with your agents via mcp, then build your own apps and plugins on top of it
everything lives in repos you can share and extend. a bit like packages for knowledge work
4
2
30
1,334
okay hear me out: an agent framework that operates on top of composable repos, with a type engine that reads them as one typed graph
thats the substrate
repos are like packages for knowledge and the capabilities for working with it
you can package your knowledge together with the skills and mcp tools your agents use to work on it
the ui components also live there
the app is a shell that renders views on top of your substrate
and those views are themselves defined in the substrate, through the same type system
so you and your agent can work on your knowledge AND build the environment you use to work with it
say youre building your company brain in there
your customer notes and deal history already live in the graph
you dont want to pay for a separate crm system when the data is already there
so you and your agent build a crm on top of that data, right inside the same app
e.g. a view of your sales pipeline, with each deal linked to the customer and the conversations behind it
you can also share that view together with the types and agent workflows it needs to let them use your setup (but not necessarily your data)
so someone else brings the repo into their workspace and uses it with their own company data
technical overview of @arsumbrisai
- many repos of md + yaml → one typed graph
- type refinements, cross-repo types
- lists, tuples, inline vs [[linked::repo]] refs
- abstract, sealed, unions, intersections
- typed md bodies, location rules, meta
- au-host: views as typed projections
- au-mcp: tools/hooks/skills as typed plugins
4
4
25
1,940
btw heres the github link if you are curious (or at least the entry point for the 40+ repos haha): github.com/arsumbris/arsumbr…
2
188
i dont think post-training will make md files redundant
i want to be able to read what my agent is working with and check if it makes sense (legibility)
model weights are a black box. you cant read through them and check what the model has learned
think about all the decisions in a company and how they depend on each other
when you work with other people (or agents), you want those decisions and relationships written down somewhere so you can check them and agree on what they mean
of course some procedural skills may become less useful. but you still need a surface where you map out what you know in a way thats verifiable for you
pretty sure markdown will stay because its a format that both humans and agents can read, inspect and edit
we just need to stop writing slop into md files
context needs to be highly curated and checked against your actual expectations and understanding
linked and TYPED markdown files let you model your knowledge / understanding in a graph
damn it, just stop writing slop and start modeling your knowledge
4
15
2,322
Heinrich retweeted
a 22-minute technical deep dive into our agent framework and knowledge IDE
technical overview of @arsumbrisai
- many repos of md + yaml → one typed graph
- type refinements, cross-repo types
- lists, tuples, inline vs [[linked::repo]] refs
- abstract, sealed, unions, intersections
- typed md bodies, location rules, meta
- au-host: views as typed projections
- au-mcp: tools/hooks/skills as typed plugins
1
15
2,212
ehm guys... there is a 3d capybara pet in my IDE
@feriederich said everybody should design their own pet, what the hell is going on?
13
1,464
Heinrich retweeted
levels of building ai knowledge systems:
lvl 1: markdown files + wikilinks
lvl 2: types + checks for your domain
lvl 3: teach agents your workflows as skills
lvl 4: build custom views/apps on top of your knowledge
lvl 5: build a missing mcp tool
lvl 6: package knowledge, mcp tools, skills AND ui components in one repo anyone can use in @arsumbrisai
16
3
242
21,380
apps should be different ways of working with the same substrate
(and they should be part of it too)
1
6
1,221
each layer of an ai knowledge system should solve a problem
( i use research as an example )
problem: every session starts with you explaining the project again
→ keep the context in markdown files + wikilinks
problem: every new paper needs the same instructions
→ turn your process for extracting and connecting claims into a skill
problem: you save conclusions but lose the reasoning behind them
→ model claims with source passages and typed relationships
problem: you keep reopening the same papers to compare their findings
→ build a comparison view with their methods, datasets and results side by side
problem: ten papers support a claim, but you cant tell how much of their evidence overlaps
→ build an mcp tool that runs a script to traverse the evidence graph and flag shared studies or datasets
problem: someone else wants to work this way
→ package knowledge, types, skills, mcp tools AND ui components in one repo they can use and extend in ars umbris
levels of building ai knowledge systems:
lvl 1: markdown files + wikilinks
lvl 2: types + checks for your domain
lvl 3: teach agents your workflows as skills
lvl 4: build custom views/apps on top of your knowledge
lvl 5: build a missing mcp tool
lvl 6: package knowledge, mcp tools, skills AND ui components in one repo anyone can use in @arsumbrisai
5
33
9,725