Show HN: A Karpathy-style LLM wiki your agents maintain (Markdown and Git)
Developers are experimenting with “LLM wikis” – markdown- and git-based knowledge bases that AI agents can read from and write to, inspired by Andrej Karpathy’s idea of treating a repo as an evolving memory for agents. Commenters debate durability (plain-text markdown, git history), retrieval strategies (BM25 vs vectors, routing queries by context), and how to constrain agents so they don’t generate low-quality “slop” or overwhelm the system with noise. Many argue these tools are most valuable when agents handle structure, indexing, and routine updates while humans curate promotions and critical content, especially for team-wide rather than purely personal use.
Markdown, Git, and Durability
- Markdown is praised as open, simple, widely supported, and likely to remain readable long term.
- Git versioning is seen as a natural fit for tracking evolving agent-written artifacts.
- Some question whether markdown itself improves LLM performance or is mainly about distribution and tooling.
Obsidian vs Dedicated Agent Wiki
- Several suggest “just use an Obsidian vault + plugins.”
- Project authors argue Obsidian is single‑user–centric, lacks promotion workflows and machine-facing APIs (MCP tools), but can still act as a read-only or parallel editor on the same markdown tree.
Retrieval Strategy: BM25, Vectors, Indexing
- Many approve the “BM25-first” design over defaulting to vector databases.
- Discussion on routing: text length/shape is a weak signal; using the agent’s task context may be better for choosing between exact-match vs narrative retrieval.
- Others propose simple indices or TOCs; counterpoint is that cascaded filtering reduces context noise and makes reranking feasible.
Quality of Agent-Generated Wikis
- Strong skepticism that “teams of agents” mostly produce low-quality “slop.”
- Others report positive experiences when agents operate over a curated, git-based knowledge base, improving coordination across tools and repos.
- Cited research suggests fully LLM-maintained docs can degrade quality vs human-maintained; hybrid setups with human curation work better.
Note-Taking Philosophy and Noise
- Some reject automated note-taking entirely: the value is in humans building their own mental models.
- Others use agents for structuring, tagging, and refactoring notes, while keeping humans responsible for actual understanding.
- Concern that AI makes it too easy to generate mountains of text nobody reads.
Governance, Promotion, and Knowledge Decay
- Multiple comments stress separating “capture” from “promotion”: agents can draft freely, but trusted entries need human review or multi-agent agreement.
- Worries about confidently wrong entries compounding over time and being re-cited.
- Questions raised about missing features like temporal vs atemporal memory, snapshots, rollbacks, and explicit handling of business rules.
Deployment, Privacy, and Provider Support
- Current model is local use with git, without pushing to public hosts; some want easy self-hosted, multi-user setups.
- OpenAI-compatible endpoints (including local or alternative providers) are reportedly supported via an intermediate runtime.
Ecosystem, Overlap, and Hype
- Noted that multiple LLM-wiki systems hit the front page in a day; some see duplication and wish for collaboration.
- Mixed reactions to the product’s playful branding: some find it slick; others see it as satire or fad-chasing, while maintainers emphasize prior serious CRM/context-infra work underpinning it.