Kevin-Wiki

Kevin's personal LLM wiki and second-brain agent harness: a flat-Markdown, agent-authored knowledge system where the wiki is the codebase, agents are the maintainers, and every runtime inherits the same brain.

What it is

Kevin-Wiki is the operating system behind Kevin's agent work. It stores durable knowledge, preferences, skills, routing maps, automations, project history, people, research, design references, and career material in readable Markdown. The central rule is that Kevin curates sources and asks questions while agents write, cross-reference, maintain, and verify the wiki. Source: AGENTS.md; raw/github/readmes-2026-06-25/Kevin-Wiki.md

The project follows the Karpathy LLM Wiki pattern: no database for the core knowledge layer, flat Markdown in wiki/, and local search through qmd rather than a hosted vector DB. The schema is meant to be readable by agents before it is optimized for a UI. Source: raw/github/readmes-2026-06-25/Kevin-Wiki.md

The current framing is stronger: Kevin-Wiki is my LLM wiki for an agent harness that acts as a second brain. It is the canonical brain and harness preset for Cursor, Claude Code, Codex, Hermes, OpenClaw, and future runtimes. The runtime may change; the identity spine, compiled wiki, skill source tree, routing maps, automations, generated packs, and proof logs remain the shared source of truth. Source: User request, 2026-07-01; LLM Wiki Agent Harness

Matt Pocock's public agent-managed Karpathy-style wiki experiment is a useful peer signal for the same thesis: live source ingests across communication surfaces become the knowledge base/environment for future loops, and the payoff is an agent that can reason with the whole business context instead of isolated prompts. Kevin-Wiki already implements that direction through source capture, compiled pages, qmd retrieval, skills, automations, and logs. Source: X/@mattpocockuk, 2026-07-01; Capture Ingest Protocol

System contract

Layer Role
raw/ Immutable source captures: X bookmarks, articles, AI sessions, Google exports, meeting notes, and generated snapshots.
wiki/ Compiled knowledge: pages with frontmatter, citations, wikilinks, typed relationships, and append-only timelines.
AGENTS.md The schema and operational contract for Codex, Claude Code, Cursor, and future agents.
skills/ Executable SKILL.md procedures that agents load before repeatable tasks.
automations/ Recurring jobs that refresh sources, run radars, compile material, and maintain the system.
ui/ Next.js wiki reader with native Markdown rendering, diagrams, search, graph views, and article affordances.

Source: AGENTS.md; raw/github/readmes-2026-06-25/Kevin-Wiki.md

Product thesis

The wiki is a proof that agents become more useful when they inherit durable context and write back to it. Chat alone makes the user carry state. Kevin-Wiki makes the system carry state: boot files, routing hubs, source captures, project pages, skills, and doctors all make the next conversation cheaper and less ambiguous. Source: The Brain-Agent Loop; No One-Off Work; AGENTS.md

The no-one-off rule is the design center. If Kevin asks for a recurring workflow twice, the system should turn it into a skill, script, doctor, or automation instead of leaving it as a memory of a one-time chat. Source: AGENTS.md; No One-Off Work

The Hermes scan sharpened the boundary: a useful agent system needs separate surfaces for compact identity, compiled knowledge, procedural skills, tool authority, context files, cron, memory/search, and proof. Kevin-Wiki owns those surfaces as durable source and projects them into runtimes; Hermes owns the live daemon, gateway, terminal backends, and cron execution when it is the selected runtime. Source: Hermes architecture docs, 2026-07-01; Hermes Harness (Nous Research); LLM Wiki Agent Harness

Current shape

The README snapshot describes the wiki as a persistent, compounding knowledge base maintained by agents, with qmd for on-device BM25/vector search, scripts/build-index.ts for catalogs/backlinks, scripts/check-freshness.ts for stale automation/sync detection, and a Next.js viewer for browsing. Source: raw/github/readmes-2026-06-25/Kevin-Wiki.md

As of the 2026-06-25 session freshness check, the repository had 1363 pages. The GitHub description and README use the broader public positioning of "1k+ pages across 25 categories." Treat exact counts as generated facts that change after every ingest. Source: npx tsx scripts/check-freshness.ts, 2026-06-25; raw/github/kevin-repos-2026-06-25.json

Why it matters in Kevin's portfolio

Kevin-Wiki is the strongest proof that Kevin is not only building agent infrastructure at work; he is also using agents as a personal operating layer. It connects Loop (skills that update), Agent Machines (persistent workers), Sigil UI (agent-editable design), and Career Profile (artifact-led positioning) into one maintained system. Source: User profile synthesis, 2026-06-25

It should appear in project/profile contexts as: personal LLM wiki and second-brain agent harness; 1k+ agent-authored pages; flat Markdown + qmd; cross-harness skills, routing, automations, doctors, generated packs, and Next.js reader; "agents write it, Kevin reads it." Source: raw/github/readmes-2026-06-25/Kevin-Wiki.md; AGENTS.md; User request, 2026-07-01

Current Evidence (2026-07-01)

This page was refreshed against Kevin's live GitHub inventory, pinned repositories, PortfolioMon cards, README scrape, and downloaded portfolio screenshots. Treat this section as the current evidence anchor; deeper interpretation stays in the surrounding compiled truth.

GitHub repo Visibility Language Stars Last push Homepage
Kevin-Liu-01/Kevin-Wiki private Python 1 2026-07-01 site

Timeline