Show HN: A Karpathy-style LLM wiki your agents maintain (Markdown and Git)
This project, WUPHF, introduces a 'Karpathy-style' wiki for AI agents, allowing them to collaboratively build a persistent knowledge base using humble Markdown and Git. It stands out for its deliberate choice of a minimalist tech stack (Markdown, Git, Bleve, SQLite) over complex vector databases, aiming for durability and local control. Hacker News found it compelling for its pragmatic approach to AI agent memory and the thoughtful design choices behind its 'shared brain' concept.
The Lowdown
WUPHF (Workstation for Ubiquitous Pervasive Humanoid Functions) presents an innovative "office" environment for AI agents, designed to foster collaboration and shared knowledge. Inspired by Karpathy's ideas on LLM-native knowledge substrates, WUPHF aims to create a persistent memory layer where context compounds across sessions, allowing agents to read from and write into a shared wiki.
- Agent Notebooks & Shared Wiki: Each agent gets a private Markdown notebook, while the team shares a central wiki. Knowledge flows from individual observations to canonical wiki entries via a human or agent-driven promotion process.
- Durability & Version Control: The wiki leverages Markdown files and Git for source-of-truth storage, ensuring human readability, version history, and portability. Changes are committed under distinct Git identities like "Pam the Archivist" for provenance.
- Efficient Retrieval: It uses Bleve for BM25 indexing and SQLite for structured metadata. A heuristic classifier routes short queries to BM25 and more narrative ones to a cited-answer loop, prioritizing efficiency over immediate reliance on vector databases.
- Context & Consistency: Features like [[Wikilinks]] with broken-link detection and a daily lint cron for contradictions, stale entries, and broken links help maintain knowledge quality and consistency. Canonical IDs ensure stable references.
- Modularity & Open Source: WUPHF is open source (MIT), self-hosted, and allows integration with various LLM CLIs (Claude Code, Codex, OpenClaw). Users can adopt just the wiki layer or the full collaborative office suite.
By building on widely understood and durable technologies like Markdown and Git, WUPHF offers a refreshing, back-to-basics approach to persistent memory for AI agents, emphasizing simplicity, auditability, and local control while still achieving robust knowledge management capabilities.
The Gossip
Markdown's Might and Minimalist Memory
Commenters appreciated the choice of Markdown and Git for the knowledge base, highlighting its durability, human readability, and versioning benefits. There was also a natural query about why not use existing tools like Obsidian, which the author addressed in the original post by implying this project offers a more programmatic and agent-native approach beyond a simple plugin.
Retrieval Rationale & BM25 Bet
The project's decision to prioritize BM25 for retrieval, with an 85% recall@20 internal ship gate, sparked interest. Questions arose about the heuristic classifier that distinguishes between 'short lookups' and 'narrative queries' for routing, particularly how it handles ambiguous agent-generated queries and potential impacts on latency and cost if complex queries are misrouted to the more expensive cited-answer loop.
Karpathy's Kinship & Criticality
The project's explicit reference to 'Karpathy-style' LLM wikis prompted discussion around Andrej Karpathy's influence in the AI community. Some commenters provided context by linking to Karpathy's original post, while others expressed a general skepticism or 'dislike' for what they perceived as 'Karpathy's fanatism towards LLMs,' albeit without detailed elaboration on the reasons behind this sentiment.