Luis

building cool stuff

[Yousef @yousefhll]

Introducing Vendo (YC S26), the customization layer that lets your users build their own features and micro-apps on top of your product

For 50 years, software has been shipped one way for everyone. That era is over.

With Vendo, every user shapes your product into exactly what they need. This is what software of the agentic era will look like. This is Vendo.

One command to install. Open source.

@runvendo @nourzahzah

[Luis @luisf_mc]

Nice morning @cursor_ai

[Luis @luisf_mc]

i was trying codex’s and claude’s internal browsers, and it’s impressive how much better gpt-5.6 sol is at manipulating codex’s browser than fable 5 is at manipulating claude’s

fable still outperforms sol in some cases, but openai did an amazing job here

[Luis @luisf_mc]

one of the best things of sol is the inability of letting it go. i love it

[Luis @luisf_mc]

damn

[Luis @luisf_mc]

anthropic and openai fighting for our attention means users keep winning

[ClaudeDevs @ClaudeDevs]

We've reset 5-hour and weekly rate limits for all users.

[Luis @luisf_mc]

markdown is doing more work here than it looks

inspectable, diffable, editable. the memory has an interface humans can actually debug

[shyam @shyamsundar]

Over the last month, I built a long-term memory system for my AI agent.

Not a folder of notes.

Not a vector database with random chunks.

A structured knowledge system that my agent can actually query, reason over, and improve.

The stack combines three ideas:

1. Karpathy’s LLM Wiki pattern

Raw information should not stay raw.

Meetings, notes, links, transcripts, decisions, and documents get compiled into a clean, persistent, human-readable Markdown wiki.

Customers, products, decisions, experiments, and concepts become dedicated pages with citations and backlinks.

The goal is not summarization.

The goal is compounding knowledge.

2. GBrain-style graph retrieval

Once the knowledge lives in structured Markdown, it can become a graph.

Vector search is useful for finding semantically similar text.

But company knowledge is often relational.

“What did customers say about onboarding?” is a semantic question.

“Which customer pain led to which feature, decision, roadmap item, and launch note?” is a graph question.

That requires entities, relationships, timelines, backlinks, and traversal.

My current system has:

1,257 pages
8,778 chunks
951 indexed Markdown documents
semantic search
graph retrieval
typed relationships
timelines
citations
backlinks

3. Hermes as the agent layer

The final step is giving the agent access to this memory.

Instead of starting from zero every session, Hermes can query the wiki and retrieval layers to find old decisions, surface patterns, detect contradictions, update pages, and improve its own workflows over time.

The architecture looks like this:

raw inputs
→ meetings, notes, docs, links, transcripts

markdown wiki
→ structured pages, citations, backlinks

gbrain
→ chunks, entities, graph edges, timelines

qmd/vector index
→ semantic retrieval over markdown

hermes agent
→ long-term memory, synthesis, self-improvement

The biggest lesson:

Don’t start with a vector database.

Start with a source of truth humans can read and edit.

Then add search.

Then add graph structure.

Then give agents access to it.

Most company knowledge bases are digital graveyards.

I think the next generation will look more like operating memory:

human-readable, machine-queryable, graph-connected, and agent-operated.

[Luis @luisf_mc]

the hard evals start when two real sources disagree and only one is current

[Luis @luisf_mc]

a useful eval case:

at 2:14, the agent approves a refund.
at 2:16, a payment dispute arrives.

what should happen before the next action?

[Luis @luisf_mc]

@LunaBitar congrats!

[Luis @luisf_mc]

this is soooooo nice

[Jameson Detweiler @jamesoniam]

The next great American car is a robot.

Talks. Parks itself. More customized than your coffee.

Meet Chip (@drivewithchip). You can reserve today. Starting at $15,000.

[Luis @luisf_mc]

quickest path to burnout

[Han Wang @handotdev]

https://t.co/uFkkNhc2Dx

[Luis @luisf_mc]

what should an agent do when two real sources disagree?

old crm field
new support ticket
stale policy doc

“pick the most relevant chunk” is not an answer

[Luis @luisf_mc]

@scale_AI trust is doing a lot of work here

for clinicians it usually means “will this hold up when the chart is messy and the patient is in front of me?”

[Luis @luisf_mc]

@lilianweng nice

feels like the hard bit is less “can it improve” and more “can we tell what changed without reading the whole trajectory by hand.

[Luis @luisf_mc]

@llama_index @TauriaP how much layout survives the trip?

tables and footnotes are usually where pdf extraction starts lying quietly.

[Zara Zhang @zarazhangrui]

People: how do you have time to do so much??

Me: I just love what I do

[Luis @luisf_mc]

it’s amazing how a gym session refreshes the mind before coming back to work

[Luis @luisf_mc]

most agent bugs will look boring.

old plan tier.
stale policy.
closed ticket treated as current.

then the write happens.

[Luis @luisf_mc]

me wearing headphones with nothing playing just so it’s only me and the code

[Mikhail Rogov @i_mika_el]

what’s actually playing in your
headphones when you code?

I cycle between lo-fi and total silence
depending on how broken things are 😅