Karpathy's Layered Wiki Solves LLM Context Loss

LLMs forget prior context across sessions because their knowledge is limited to the current prompt. Karpathy proposes a local "wiki" modeled after Obsidian's graph structure: interconnected markdown nodes with infinite layers of links for efficient traversal. Instead of flat files, organize knowledge hierarchically—one entry point links to sub-skills or details, enabling agents to hop through layers without token overload.

Key infrastructure: an index acts as a table of contents for quick catalog access; an append-only log provides chronological history like git commits, tracking updates without overwriting. Pull external knowledge (documents, research) into this curated base so the LLM "knows" it natively. This persists projects like plans or code, avoiding re-explanation each time. His gist details implementation: curate files, let agents traverse/update the graph.

Trade-off: manual setup with Obsidian requires downloading, configuring agents, and maintenance—hands-on work that delays production use.

Manual Workflows Waste Time on Repetition

Without persistence, workflows degrade: research competitors in Claude, save output, then three days later paste it back for a blog post—losing links to sources. A week on, newsletter overlaps blog content due to no awareness. Pasting takes 8-12 minutes per session (less than full re-research but still manual), no logging, no version control, inconsistent tone without baked-in frameworks.

Agents can't self-reference prior work, leading to redundant research (e.g., competitor pricing changes ignored) or hallucinated continuity. Flat markdown repos (like author's Consume, built on GitHub) lack depth—code manipulates files but misses graph traversal and shared history.

Cabinet Productizes the Wiki for Human-Agent Teams

Cabinet (1,000+ GitHub stars, open-source, npx runnable, Mac app soon) packages Karpathy's idea into a shareable "cabinet": folders with markdown, sheets, HTML apps. Agents query/update the internal repo automatically—research agent scans competitor list weekly, detects changes (pricing/launches), increments versions, logs to changelog, flags content agent for review.

Automation loop: Agents maintain source-of-truth databases, quantify "significant" changes to avoid noise, enable reversion via logs. Humans interact via apps—HTML/JS dashboards for visualization (e.g., competitor profiles), editable without databases (updates sync to markdown).

Share cabinets to transfer full project context between people/agents, like zipping a workplace. Beats Paperclip's shallow agent.md templates by supporting multi-file depth (HTML, sheets). For consultants, hold client knowledge in one portable GUI; organizations get middleware for AI integration—readable changelogs explain agent actions.

Run locally: npx for instant start, no cloud. Author plans to migrate Consume from GitHub to cabinets for better portability. Outcome: continuous, live documentation that scales knowledge work, blending human curation with agent maintenance.