Beyond RAG: Implementing Karpathy's Persistent LLM Wiki Pattern with Hjarni

Karpathy’s LLM Wiki is right. I just didn’t want to run it locally.

Andrej Karpathy’s “LLM Wiki” gist proposes a persistent markdown-based knowledge store maintained by LLM agents rather than stateless RAG systems. This pattern shifts the synthesis tax from query time to ingestion, allowing knowledge to compound through automated cross-references and contradiction flagging.

Why This Matters

Standard RAG architectures suffer from a synthesis tax where knowledge is re-derived on every query, leading to inconsistent outputs. While local setups using Obsidian and Claude Code provide a foundation, they create technical silos restricted to single machines and specific clients. Moving to a hosted Model Context Protocol (MCP) interface allows for a “living brain” accessible across mobile apps, IDEs like Cursor, and various LLM clients without the friction of manual synchronization, solving the maintenance bottleneck that typically causes human-managed wikis to fail.

Key Insights

• Persistent synthesis vs RAG: Instead of re-deriving knowledge per query, an LLM agent incrementally maintains markdown files to flag contradictions and build cross-references.
• The Bookkeeping Bottleneck: Human maintenance of wikis fails because upkeep costs grow faster than value; LLMs automate the updating of multiple pages per new source source.
• MCP for Unified Context: Hjarni uses the Model Context Protocol to expose hosted notes, tags, and links to any client including Claude, ChatGPT, and Cursor.
• Vannevar Bush’s Memex: The vision of an associative knowledge store is realized through LLMs performing the maintenance tasks that were historically impossible for humans to sustain.

Practical Applications

• Use case: Developers using Cursor for coding and the Claude mobile app for note capture can access and edit the same structured knowledge graph via Hjarni.
• Pitfall: Relying on local-only Obsidian vaults prevents knowledge access across multiple LLM clients or on mobile devices, creating an ‘information island.’
• Use case: Engineering teams using Hjarni Pro can allow multiple humans and their respective LLM agents to operate out of a single shared knowledge base.
• Pitfall: Abandoning structured documentation because the friction of manual updates grows faster than the value of the knowledge stored.

References:

• https://dev.to/hjarni/karpathys-llm-wiki-is-right-i-just-didnt-want-to-run-it-locally-170m