Join the discussion
Write your take first — we'll ask for email only when you're ready to publish.
- Hacker News
- nice :)
> love to hear how you did it
this is ours we built for Hedgy https://setoku.com
our approach was to build a data lake that sucks company data into clickhouse and staple that to a knowledge store. this way the brain has a stream of live facts and builds knowledge around it. we gave up on trying to make the knowledge store human-readable -- i totally think there could be something there, but for now we just care about enhancing the agent you're using. it makes my claude code very good at debugging and gives everyone a way to vibecode dashboards and small internal tools with real data.
i also run a personal instance for my wife and I that sucks in monarch money and gmail. mostly use it to chat through big money moves.
by rgbrgb - How has traction been so far? And are there any valuable cross team/agent use cases you actually have experienced or are they theoretical?
I run https://smalldocs.org, which focuses on making Markdown a highly expressive space for agents (think slides, spreadsheets, charts, mermaid diagrams, etc.). I have built a local library which lives in the browser but reflects your local file system - explainer here: https://smalldocs.org/connect?return=%2Flibrary, demo here: https://smalldocs.org/library?demo=1. Which is pretty useful when you want to find local Markdown files.
However I am just about to introduce very similar functionality to yours - a cloud + agent access. Right now I don't know if this will actually be useful, but I think it will likely be something I can charge for. I have a few hundred developers that use SmallDocs on a weekly basis (https://smalldocs.org/analytics), but right now it's just a free local-first tool.
Code available here: https://github.com/espressoplease/smalldocs
by FailMore - I think the central question for a such memory system is whether we or the agents can find the relevant information and how to organize these data as changes continues to come in. Would we miss something in the retrieval process? How do we organize the information so they stay actual and correct without piling up the garbage? Of course we can continue to concatenate the data and tag them with version and date, but then we have to face the problem of extracting the relevant information in a short time. If we delegate that problem to a LLM, long context retrieval performance will degrade and the cost will explode.
That is the reason why we condense the information in the first place. Forgetting + Synthesizing are the necessary parts of learning and basically with memory + smart retrieval we want to build a learning system.
by sinuhe69 - I use the LLM-wiki pattern for a structured directory of topic folders of .md files, and made it also compatible with the Open Knowledge Format [1].
My agent (Hermes) responds to a made-up command "vaultize this doc/link/text, etc" to add new .md files in the right format in the right place. The agent does a pretty good job of maintaining the index.md file, cross-links, etc.
A Quartz website builder creates a static site on my server, each MD file is a web page, and rebuilds when a new file is added [2]. This setup gives me a useful knowledge base with a great UX - across all the devices on my Tailnet, no need to run Obsidian Sync, or Syncthing.
This system is simple and works well, but i think it could maybe benefit from a memory system like Honcho to make it more effective as it scales.
[1] https://cloud.google.com/blog/products/data-analytics/how-th...
by fallinditch - I recently built our knowledge management system based (loosely) on the Karpathy LLM-Wiki model.
I'm not sure I understand why it is so important to be multi-agent, or for each agent to define who they are and what changes they made.
Our process uses git for tracking, so much of these details are captured natively in the merge.
Our process is
Drop a change, or something to document in an "inbox". Agent makes a branch, processes the inbox materials, updates the documents that need to be updated, updates timelines, makes a note of others on the team that may need to be notified of changes.
Creator reviews the PR and ensures the changes match how they want the document to be updated. Anyone that needs to be updated is now a reviewer on the PR.
I've just started rolling this out to the rest of the team, thankfully we are an engineering heavy organization so everyone is comfortable in Git.
The system keeps a running audit log, input is never deleted, it's stored, documents are updated, people are notified, etc. etc.
I'm just not sure what the service is offering, and how being multi-agent is solving a problem, but maybe I just haven't hit that problem yet.
by pedalpete - Almost all of this is stuff I have indeed "frankenstein[ed]" for myself, so consider this comment a +1 on market fit, there!
That also gives me a reason to pause, tho; the pitch in general is as solid as it can be on a site with markdown turned off (why, lord, why), but as a format minutiae megafan, I was left a little dissapointed. Where do you/OzBrain stand on Markdown formats? Could I use Sphinx with this, in rST and/or native MyST? Can it generate plain PDFs, fancy PDFs, or even animated static sites? etc. etc. etc. Not trying to gotcha, just curious to hear your thoughts & dreams on the topic!
It seems like some subculture(s) of SWE/SV/YC/AI has landed on obsidian-ish markdown with lots of wikilinks as the presumed default, which makes sense. So I'm assuming it's the same here. But also, your 'OzBrain vs. Obsidian' page does describe one difference as 'Markdown export anytime' vs. 'Markdown on disk' -- presumably that's just a hedge about hosting paradigm rather than a comment on the persistent format?
P.S. You're likely aware but there's at least one other company using Oz -- Warp's coding agent. Have you considered renaming this to something unimpeachable like DeepReasoningBrain? ;)
P.P.S. Holy hell your `eng-flow` thing is incredible. Maybe I'm behind the times, but... I mean, has anyone else processed how close we are to Minority Report and Iron Man?!
P.P.P.S. Is any part of that/this OS?
by bbor - I have a folder called reports, plans, and code-reviews in each repo. I put my md files for agents there, and voila they're in the cloud along with my source code in git. I just talk to my local agent about these files and it finds things using grep and whatever. Done. No mcp or special server needed.
I've been pitched products like ozbrain before, but I've failed to see the need over what I already have. Seems like more complication for no gain to me.
Am I missing something?
by Sammi - Do you have a solution for degradation in accuracy when compiling larger amounts of llm-produced text?
I am also building LLM knowledge/memory systems and I've been surprised how bad LLMs are, even SOTA models, at summarizing non-trivial input batches of text. They get things wrong, distort the underlying meaning or data, etc.
by gavinboston