ES version is available. Content is displayed in original English for accuracy.
It runs locally in ~/.wuphf/wiki/ and you can git clone it out if you want to take your knowledge with you.
The shape is the one Karpathy has been circling for a while: an LLM-native knowledge substrate that agents both read from and write into, so context compounds across sessions rather than getting re-pasted every morning. Most implementations of that idea land on Postgres, pgvector, Neo4j, Kafka, and a dashboard.
I wanted to go back to the basics and see how far markdown + git could go before I added anything heavier.
What it does: -> Each agent gets a private notebook at agents/{slug}/notebook/.md, plus access to a shared team wiki at team/.
-> Draft-to-wiki promotion flow. Notebook entries are reviewed (agent or human) and promoted to the canonical wiki with a back-link. A small state machine drives expiry and auto-archive.
-> Per-entity fact log: append-only JSONL at team/entities/{kind}-{slug}.facts.jsonl. A synthesis worker rebuilds the entity brief every N facts. Commits land under a distinct "Pam the Archivist" git identity so provenance is visible in git log.
-> [[Wikilinks]] with broken-link detection rendered in red.
-> Daily lint cron for contradictions, stale entries, and broken wikilinks.
-> /lookup slash command plus an MCP tool for cited retrieval. A heuristic classifier routes short lookups to BM25 and narrative queries to a cited-answer loop.
Substrate choices: Markdown for durability. The wiki outlives the runtime, and a user can walk away with every byte. Bleve for BM25. SQLite for structured metadata (facts, entities, edges, redirects, and supersedes). No vectors yet. The current benchmark (500 artifacts, 50 queries) clears 85% recall@20 on BM25 alone, which is the internal ship gate. sqlite-vec is the pre-committed fallback if a query class drops below that.
Canonical IDs are first-class. Fact IDs are deterministic and include sentence offset. Canonical slugs are assigned once, merged via redirect stubs, and never renamed. A rebuild is logically identical, not byte-identical.
Known limits: -> Recall tuning is ongoing. 85% on the benchmark is not a universal guarantee.
-> Synthesis quality is bounded by agent observation quality. Garbage facts in, garbage briefs out. The lint pass helps. It is not a judgment engine.
-> Single-office scope today. No cross-office federation.
Demo. 5-minute terminal walkthrough that records five facts, fires synthesis, shells out to the user's LLM CLI, and commits the result under Pam's identity: https://asciinema.org/a/vUvjJsB5vtUQQ4Eb
Script lives at ./scripts/demo-entity-synthesis.sh.
Context. The wiki ships as part of WUPHF, an open source collaborative office for AI agents like Claude Code, Codex, OpenClaw, and local LLMs via OpenCode. MIT, self-hosted, bring-your-own keys. You do not have to use the full office to use the wiki layer. If you already have an agent setup, point WUPHF at it and the wiki attaches.
Source: https://github.com/nex-crm/wuphf
Install: npx wuphf@latest
Happy to go deep on the substrate tradeoffs, the promotion-flow state machine, the BM25-first retrieval bet, or the canonical-ID stability rules. Also happy to take "why not an Obsidian vault with a plugin" as a fair question.
Discussion (112 Comments)Read Original on HackerNews
The whole point of taking notes for me is to read a source critically, fit it in my mental model, and then document that. Then sometimes I look it up for the details. But for me the shaping of the mental model is what counts
Highly debatable whether it’s possible to create anything truly valuable (valuable for the owner of the product that is) with this approach, though. I’m not convinced that it will ever be possible to create valuable products from just a prompt and an agent harness. At that point, the product itself can be (re)created by anyone, product development has been commodified, and the only thing of value is tokens.
My hypothesis is that “do things that don’t scale”[0] will still apply well into the future, but the “things that don’t scale” will change.
All that said, I’ve finally started using Obsidian after setting up some skills for note taking, researching, linking, splitting, and restructuring the knowledge base. I’ve never been able to spend time on keeping it structured, but I now have a digital secretary that can do all of the work I’m too lazy to do. I can just jot down random thoughts and ideas, and the agent helps me structure it, ask follow-up questions, relate it to other ongoing work, and so on. I’m still putting in the work of reading sources and building a mental model, but I’m also getting high-quality notes almost for free.
[0]: https://www.paulgraham.com/ds.html
I actually think that the harnesses which do end up building products, the harness will be the product.
As an example, I have a harness which I have my entire team use consistently. The harness is designed for one thing: to get the results I get with less nuanced understanding of why I get it.
Mind you, most of my team members are non-technical, or at least would be considered non-technical, two years ago.
These days, I spend most of my time fine-tuning the harness. What that gives me is a team which is producing at 5x their capacity from three months ago, and I get easier to review, more robust pull requests that I have more confidence in merging.
It's still a far cry from automating the entire process. I still think humans need to give the outcomes to even the harnesses to produce the results.
I think your take is right. This isn't going to help with the internalization of knowledge that note taking will get you. I do think that there is some value in the way we've set up blueprints of agents if you haven't set up a business before to either teach about role functions in a business or get a head start on business that doesn't create something new. At the very least it's a quick setup to getting to experiment.
To the part about note taking (and disclosure) - we are working on a context graph product that lessens the work of reading sources, especially over time and breath to help with a lot of the structure you've mentioned.
Say more?
It circles back to the question, is this unimportant enough for me to delegate it to a LLM that might get it wrong? If the answer is yes, why even do it to begin with. If the answer is no, you have to do it manually.
I personally though, see value in this type of automation. Stuff like tag categorization, indexing, that otherwise would've been lost seems like a good fit for LLMs. Whether or not they're an ideal solution and something else like a search engine would've been a better fit, is a different question.
Most things do not need to end up in your notes, and LLMs add too much noise, one that you likely never personally verify/filter out at all.
JA Westenberg made a good video essay about it a few days ago:
https://youtube.com/watch?v=3E00ZNdFbEk
I think the sweet spot is human curation of these documents, but unsupervised management is never the answer, especially if you don’t consciously think about debt / drift in these.
I'd love to see other sources that seek to academically understand how LLM's use context, specifically ones using modern frontier models.
My takeaway from these CLAUDE.md/AGENTS.md efforts isn't that agents can't maintain any form of context at all, rather, that bloated CLAUDE.md files filled with data that agents can gather on the spot very quickly are counter-productive.
For information which cannot be gathered on the spot quickly, clearly (to me) context helps improve quality, and in my experience, having AI summarize some key information in a thread and write to a file, and organize that, has been helpful and useful.
I've been coding like this lately: if I'm too lazy to review a new non-critical section/unit tests, I'll mark it as `// SLOP`; later, if I have to, I'll go through the entire thing, and unmark
shitty tests are better than no tests, as long as you your expectations are low enough
it's neat, i can create a new sprite/whatever, point claude at the root, and tell it to setup zswap and it will know exactly how to do so in that environment. if something changes, and there's some fiddling to make it work, i can ask it to write a report and send it in to fold into the existing docs.
This LLM stuff at least has some real possibilities and value, and is very fun tech to learn about and play with.
I long ago accepted that there’s money to be made, as long as it’s not unethical, then get involved. Can build cool things that do have value, while enjoying the VC/PE money sloshing around
not interested in a job at Anthropic as Principal Engineer (i used to be a HubSpot Product Manager with a healthy income, much better than what i am making now, or for the next few years).
took multiple bets and did iterations because we talked to customers and kept evolving while our old competition is still building an AI CRM "in stealth".
been around enough to know waves don't matter but there is still value behind those waves worth extracting away.
fun and usable can co-exist and we are trying the best to prove that. also, we have an amazing designer who never worked at big tech and has no accolades, but man got taste.
It is a third llm wiki on front page in 24 hours! Obviously it is a hot topic. I have my own horse in that race - so I might not be objective - but I've compiled a wishlist for these system: https://zby.github.io/commonplace/notes/designing-agent-memo...
I wish there was a chance for collaboration - everybody coding their own system seems like a lot of effort duplication.
But the reviews are written automatically - here are the instructions: https://github.com/zby/commonplace/blob/main/kb/agent-memory...
Overall the knowledgebase is a mixture of these. I have this disclaimer on the first page:
This KB is itself agent-operated: a human directs the inquiry, AI agents draft, connect, and maintain the notes. The framework for building knowledge bases is documented using that framework.
I hope it is enough - I've seen many people get angry with publishing LLM generated work.
full disclosure: we started as a context infra company (nex.ai) from long long before Karpathy even came up with the LLM wiki idea, and have barely exposed any of that stuff to WUPHF but starting to open some of that now. glad to see the concerns in the comparison are things that our context infra already built for.
still, happy to collab & share learnings, and of course avoid duplication.
https://x.com/karpathy/status/2039805659525644595
https://xcancel.com/karpathy/status/2039805659525644595
I would be fine with my agent saying “hey, we hit a wall here, here’s the question posted on SO, I flagged to come back to it later once we have an answer”
One example experiment I had was seeing if I could get a llm to build its own knowledge wiki where I would paste a few links and have it go do some research on whatever the subject was, then distill what it finds into specific wiki pages with links to other pages or the source refs. It looked good until you read the actual data.
This was a few years ago though, so maybe it's worth trying again with something like opus 4.7.
[1] https://tiddlywiki.com/
For the single-file configuration you describe, there are already several LLM connectors — e.g. [1]. The appeal is exactly what you note: no dependencies, no installation, trivially archivable. A single-file agentic wiki wandering around and self-editing isfeasible today.
For something closer to Karpathy's LLM Wiki pattern, I've been working on twillm[2], which uses TiddlyWiki's Node.js configuration. That setup saves tiddlers as individual files, so you can point it at an existing Markdown vault and work alongside tools like Claude Code.
Some benefits of TiddlyWiki for this:
* Open source, so you can be confident it'll remain usable indefinitely.
* Web-based, so accessible anywhere.
* Computed views replace materialised index files. Karpathy's setup relies on an index.md that the LLM has to keep in sync as it adds notes — something LLMs are bad at, with staleness creeping in across sessions. TiddlyWiki views are live filter expressions: "tiddlers tagged concept, sorted by rating" computes its contents at render time.
* Frontmatter becomes queryable structure. Obsidian renders YAML frontmatter as boxed metadata at the top of a note. TiddlyWiki promotes frontmatter fields into first-class tiddler fields you can filter, sort, and aggregate over.
* LLM-authored applets, not just content. Beyond Markdown notes, the LLM can drop in wikitext tiddlers (.tid) that act as small interactive live views: dashboards, browse-by-tag tools, journal indexes, glossary pages.
[1] https://github.com/rimir-cc/tw-llm-connect [2] https://github.com/Jermolene/twillm
I've also experimented recently with such a project [0] with minimal dependencies and with some emphasis on staying local and in control of the agent.
It's building and organising its own sqlite database to fulfil a long running task given in a prompt while having access to a local wikipedia copy for source data.
A very minimal set of harness and tools to experiment with agent drift.
Adding image processing tool in this framework is also easy (by encoding them as base64 (details can be vibecoded by local LLMs) and passing them to llama.cpp ).
It's a useful versatile tool to have.
For example, I used to have some scripts which processed invoices and receipts in some folders, extracting amount date and vendor from them using amazon textract, then I have a ui to manually check the numbers and put the result in some csv for the accountant every year. Now I can replace the amazon textract requests by a llama.cpp model call with the appropriate prompt while still my existing invoices tools, but now with a prompt I can do a lot more creative accounting.
I have also experimented with some vibecoded variation of this code to drive a physical robot from a sequence of camera images and while it does move and reach the target in the simple cases (even though the LLM I use was never explicitly train to drive a robot), it is too slow (10s to choose the next action) for practical use. (The current no deep-learning controller I use for this robot does the vision processing loop at 20hz).
[0]https://github.com/GistNoesis/Shoggoth.db/
That means the longer an agent runs on a task, the more likely it will fail the task. Running agents like this will always fail and burn a ton of token cash in the process.
One thing that LLM agents are good at is writing their own instructions. The trick is to limit the time and thinking steps in a thinking model then evaluate, update, and run again. A good metaphor is that agents trip. Don't let them run long enough to trip. It is better to let them run twice for 5 minutes than once for 10 minutes.
Give it a few weeks and self-referencing agents are going to be at the top of everybody's twitter feed.
> "The LLM model's attention doesn't distinguish between "instructions I'm writing" and "instructions I'm following" -- they're both just tokens in context."
That means all these SOTA models are very capable of updating their own prompts. Update prompt. Copy entire repository in 1ms into /tmp/*. Run again. Evaluate. Update prompt. Copy entire repository ....
That is recursion, like Karpathy's autoresearch, it requires a deterministic termination condition.
Or have the prompt / agent make 5 copies of itself and solve for 5 different situations to ensure the update didn't introduce any regressions.
> reach local maximima unless external feedback is given
The agents can update themselves with human permission. So the external feedback is another agent and selection bias of a human. It is close to the right idea. I, however, am having huge success with the external feedback being the agent itself. The big difference is that a recursive agent can evaluate performance within confidence interval rather than chaos.
https://github.com/RivoLink/leaf
A lot of these tools (paperclip) don't, but for our small, agent-supported team, that would be the feature we're looking for. I.e. the agents of different marketing roles talking to each other, with clearly defined 'human in the loop' decision making.
But also would like to understand how markdown helps in durability - if I understand correctly markdown has a edge over other formats for LLMs.
Also I too am building something similar on markdown which versions with git but for a completely different use case : https://voiden.md/
[0] https://hurl.dev/
https://manfre.me/posts/2026/04/build-llm-wiki-obsidian/
Right now this is setup to be run on your machine. Git is used to do versioning but we don't push that to GitHub, nor do we keep any insight into what you have or what you're doing.
If there is long term value people are getting out of Wuphf we'll be happy to build out a hosted business/enterprise compliant version.
can also do llama.cpp server, LM Studio, LocalAI, OpenRouter, Together, Anyscale, and anything else with the OpenAI surface. i am using ollama already through OpenCode.
if you hit a specific endpoint that does not work, please file an issue.
I've been building a native Mac app "Sig" around this idea: capture has to come from you. You sit down after a meeting and articulate what happened — what was decided, who committed to what, what you actually think. That articulation is the work. The AI routes it into files. If you skip that step and scrape transcripts instead, no promotion workflow saves you. You're just contributing to the garbage in, garbage out idea.
That is Sig <https://news.ycombinator.com/item?id=47901737>
https://wuphf.team/
It looks like gastown. Using AI is like children playing with a dollhouse. If it is satire, well done.
if you want the wiki only, install WUPHF and only use the wiki layer. the promotion flow, fact log, and /lookup work without the multi-agent office around them.
1 -https://x.com/__endif/status/2039810651120705569
Every time I hear someone say "I have a team of agents", what I hear is "I'm shipping heaps of AI slop".
the team is of 4 HubSpotters who built HubSpot's largest platforms - search, nav, notifs, permissions, AI.
we are in the process of opening up large pieces of our enterprise context architecture to WUPHF and also ship the cloud enterprise version of WUPHF (https://nex.ai/new-home).
A couple of weeks ago I built a git-based knowledge base designed to run agents prompts on top of it.
I connected our company's ticketing system, wiki, GitHub, jenkins, etc, and spent several hours effectively "onboarding" the AI (I used Claude Opus 4.6). I explained where to find company policies, how developers work, how the build system operates, and how different projects relate to each other.
In practice, I treated it like onboarding a new engineer: I fed it a lot of context and had it organize everything into AI-friendly documentation (including an AGENTS.md). I barely wrote anything myself, mostly I just instructed the AI to write and update the files, while I guided the overall structure and refactored as needed.
The result was a git-based knowledge base that agents could operate on directly. Since the agent had access to multiple parts of the company, I could give high-level prompts like: investigate this bug (with not much context), produce a root cause analysis, open a ticket, fix it, and verify a build on Jenkins. I did not even need to have the repos locally, the AI would figure it out, clone them, analyze, create branches using our company policy, etc...
For me, this ended up working as a multi-project coordination layer across the company, and it worked much better than I expected.
It wasn't all smooth, though. When the AI failed at a task, I had to step in, provide more context, and let it update the documentation itself. But through incremental iterations, each failure improved the system, and its capabilities compounded very quickly.
In my experience, the difference is context. Agents without structure produce slop, but with a well-curated knowledge base and iteration, they can be useful. I was just sharing a setup that has been working for me lately.
Edit: minimal changes for clarity
tbh this won't be much useful as a personal wiki.
We have been using it as a sounding board. I think that in its current state it's actually more useful for someone to learn about how to run a business - "what does a CEO vs PM do" and/or learn about the pros/cons of running a bunch of agents at once.
I built a tiny utility like this that works very well yesterday:
https://github.com/skorokithakis/gnosis