WUPHF runs a team of AI agents (Claude Code, Codex, OpenClaw) in a shared office backed by a git-native Markdown wiki agents read and write as collective memory.
Key Takeaways
Wiki lives at ~/.wuphf/wiki/ as a local git repo with typed fact triplets, per-entity append-only logs, LLM-synthesized briefs, and a /lint suite for contradictions and orphans.
Agents write to private per-agent notebooks first; only explicit agent judgment promotes a fact to the shared wiki – nothing graduates automatically.
Fresh-session-per-turn design achieves 97% Claude API prompt cache hit rate; a 5-turn session costs $0.06 vs accumulated orchestrators growing from 124k to 484k tokens over 8 turns.
Per-agent MCP scoping keeps DM mode at 4 tools and full-office mode at 27, shrinking prompts and improving cache alignment.
Pre-1.0 with daily main movement; the README explicitly says pin forks to release tags, not main.
Hacker News Comment Review
The sharpest engineering concern is LLM sprawl: multiple builders report that unconstrained agents write so much that the wiki degrades in utility as it grows, with no clear mechanism to prevent it.
Several commenters noted this is the third agent-memory or LLM-wiki project on the HN front page within 24 hours, raising fragmentation risk – each team building their own system duplicates effort without a convergence point.
Skeptics challenge the core premise: automated note-taking removes the cognitive act of shaping a mental model, so a wiki agents fill may produce stored text with no understanding behind it.
Notable Comments
@johntash: ran similar experiments and found LLMs document relentlessly until the wiki becomes a mess; asks how WUPHF prevents unbounded growth.
@mncharity: flags TiddlyWiki (20+ years old, self-modifying single HTML file) as prior art for agentic self-editing wikis, noting Git integration is the open problem there.
