Skip to content

research.md

Latest commit

 

History

History
171 lines (136 loc) · 11.4 KB

research.md

File metadata and controls

171 lines (136 loc) · 11.4 KB
Raw

Phase 1.25 — Research Report

Date: 2026-04-08 Method: Cloned every referenced implementation from Karpathy's gist and the top related GitHub searches into .temp/ (gitignored) for side-by-side comparison.

This document is the deliverable for Phase 1.25 Research, a new phase added to the llmwiki framework (see docs/framework.md). It is the source of truth for prior-art analysis and the 10x gap that llmwiki targets.

Summary

15 reference implementations were cloned and analysed. They fall into four clusters:

Cluster What they do Examples llmwiki differentiation
Pure-markdown skills A Claude Code skill/plugin + CLAUDE.md schema; rely on the agent to do all writes kfchou/wiki-skills, Astro-Han/karpathy-llm-wiki, bashiraziz/llm-wiki-template ✅ Same base + native .jsonl → markdown + static HTML + multi-agent
Markdown-first + light Python Schema + a few Python scripts for ingest/query/lint SamurAIGPT/llm-wiki-agent, Ss1024sS/LLM-wiki, hsuanguo/llm-wiki ✅ Same shape + session-transcript adapter + beautiful static site
Obsidian-coupled Wiki lives inside an Obsidian vault; user views via Obsidian AgriciDaniel/claude-obsidian, louiswang524/llm-knowledge-base, kytmanov/obsidian-llm-wiki-local, remember-md/remember 🔀 Obsidian as one of many connectors (input adapter) — not the only view
Heavy Python / hosted Backend services, databases, hosted demos lucasastorian/llmwiki (Apache, Supabase + MCP, hosted at llmwiki.app), bitsofchris/openaugi ❌ Too heavy — violates llmwiki's stdlib-first rule
Session browsers (not wikis) Search/TUI over raw .jsonl; no wiki compilation raine/claude-history, sinzin91/search-sessions 🔀 Complementary — they search raw; llmwiki builds the wiki on top

Per-repo analysis

Pure-markdown skills

kfchou/wiki-skills

  • Shape: Claude Code plugin. 6 markdown files, 0 Python files.
  • Strength: Minimal, well-scoped. Pure schema + slash commands.
  • Gap: No static HTML output. No multi-agent support. No session-transcript adapter — input is "any markdown".
  • Lesson: The "Claude Code plugin" distribution mode is clean — ship as a plugin + a few .md files, zero runtime deps.

Astro-Han/karpathy-llm-wiki

  • Shape: "One skill" marketed as Agent Skills-compatible. 6 markdown, MIT.
  • Strength: Explicitly targets the Agent Skills ecosystem.
  • Gap: Same as above — no session-aware ingestion, no HTML rendering.
  • Lesson: There's a clean "one-skill" positioning angle (minimal cognitive footprint) that resonates.

bashiraziz/llm-wiki-template

  • Shape: Template with 26 markdown files + 3 Python scripts. MIT.
  • Strength: Template layout is clean and opinionated.
  • Gap: Template alone — no session transcripts, no HTML, no adapters.
  • Lesson: A template-first distribution is a lower-effort entry point than a full tool.

Markdown-first + light Python

SamurAIGPT/llm-wiki-agent

  • Shape: 11 markdown + 4 Python scripts. MIT.
  • Strength: The best-documented schema I've read. /wiki-ingest, /wiki-query, /wiki-lint, /wiki-graph slash commands. build_graph.py generates a vis.js knowledge graph.
  • Gap: Input is "drop markdown in raw/" — no knowledge of session transcripts. No dark mode. No search. No HTML output beyond the graph HTML.
  • Lesson: This is the closest prior art. llmwiki inherits its directory layout and slash-command naming (/wiki-*).

Ss1024sS/LLM-wiki

  • Shape: 25 markdown + 4 Python. Generates 30 files including stale-reporting scripts.
  • Strength: Explicitly covers the "AI remembers nothing in new sessions" pain point in its README.
  • Gap: No HTML viewer. No multi-agent.
  • Lesson: The stale-reporting idea is worth borrowing — llmwiki should ship /wiki-lint with "stale pages" detection baked in.

hsuanguo/llm-wiki

  • Shape: 13 markdown + 9 Python. Has a logo, more polish.
  • Strength: The most Python-heavy of the markdown-first implementations — suggests the author hit real scale issues.
  • Gap: Not session-aware.
  • Lesson: Once you reach ~10 Python files, you're building a CLI. llmwiki leans into that explicitly.

Obsidian-coupled

AgriciDaniel/claude-obsidian

  • Shape: 50 markdown files, 0 Python. The wiki itself is real content living in Obsidian.
  • Strength: Treats Obsidian as a first-class viewer. The meta/ folder contains a cover GIF — good product presentation.
  • Gap: 100% Obsidian-locked. If you don't use Obsidian, this is useless.
  • Lesson: Obsidian is a compelling viewer for many users — llmwiki should ship an Obsidian connector as an optional input/output, not the only path.

kytmanov/obsidian-llm-wiki-local

  • Shape: 2 markdown + 25 Python files. Uses a local LLM (no cloud API calls).
  • Strength: Privacy angle — everything runs locally, no OpenAI/Anthropic API calls.
  • Gap: Requires running a local model — heavy prerequisite.
  • Lesson: Privacy-first positioning resonates. llmwiki's no-telemetry + local-only rules align with this.

louiswang524/llm-knowledge-base

  • Shape: 10 markdown + 1 Python. Claude Code + Obsidian.
  • Strength: Clean coupling between Claude Code (writer) and Obsidian (viewer).
  • Gap: Assumes Obsidian is installed.
  • Lesson: "Writer/viewer split" is a useful mental model.

remember-md/remember

  • Shape: 23 markdown, 0 Python. Obsidian-compatible memory for OpenClaw + Claude Code.
  • Strength: "Tool-portable memory" framing — "one brain, every AI tool".
  • Gap: Focused on memory/decisions rather than wiki compilation.
  • Lesson: The multi-agent portability angle is strong. llmwiki's adapter pattern delivers this.

Heavy Python / hosted

lucasastorian/llmwiki

  • Shape: 1 markdown + 47 Python files. Apache 2.0. Hosted demo at llmwiki.app.
  • Strength: Has a real backend (Supabase + S3), MCP server, hosted UI. Production-grade infra.
  • Gap: Requires Supabase + S3 + Node + MCP server to self-host. Violates llmwiki's stdlib rule hard.
  • Lesson: This is the "full-stack" approach. llmwiki is explicitly the opposite — zero infra, all local.

bitsofchris/openaugi

  • Shape: 27 markdown + 48 Python files. "Your augmented knowledge base for Agentic work."
  • Strength: Covers agent-centric use cases well.
  • Gap: Very heavy.
  • Lesson: There's room for a lighter alternative — which is llmwiki's niche.

Session browsers (complementary, not competitors)

raine/claude-history

  • Shape: TUI app. 3 markdown + 0 Python.
  • Strength: "Best thing ever" user quote in the README. Clear UX.
  • Gap: Search-only, no wiki building.
  • Lesson: Users love the ability to search their session history. llmwiki's wiki is additive — it compiles search-ready content.

sinzin91/search-sessions

  • Shape: CLI binary. 13 markdown + 0 Python (Go).
  • Strength: Sub-300ms search across all sessions.
  • Gap: Same — search only.
  • Lesson: llmwiki could call out to search-sessions as an optional backend for its Cmd+K search.

Not cloned (mentioned but out of scope)

The 10x gap (feature matrix)

Feature Most existing llmwiki
Ingests .jsonl session transcripts ❌ (generic markdown only) ✅
Claude Code adapter Some ✅
Codex CLI adapter None ✅ stub (v0.2)
Multi-agent adapter pattern None ✅
Pure stdlib + markdown (no DB, no MCP, no Node) ~50% ✅
Beautiful static HTML viewer ❌ ✅ god-level UI
Global search (Cmd+K) ❌ ✅ client-side index
Syntax highlighting Rarely ✅ highlight.js (CDN)
Redaction by default ❌ ✅ username + API keys + tokens + emails
Live-session detection ❌ ✅ skips <60min old
Idempotent incremental sync Some ✅ mtime state file
Windows .bat scripts Rarely ✅
Obsidian connector Some (only input) ✅ input and output
No cloud, no telemetry, no auth Some ✅ hard rule
Build time <15s for 300 sessions Varies ✅ 9s measured

Borrowed ideas (with attribution)

  • Directory layout (raw/, wiki/, index.md, log.md, overview.md) — Karpathy's gist + SamurAIGPT/llm-wiki-agent
  • Slash commands (/wiki-ingest, /wiki-query, /wiki-lint, /wiki-graph) — SamurAIGPT/llm-wiki-agent
  • Stale-page detection — Ss1024sS/LLM-wiki
  • Local-LLM privacy angle — kytmanov/obsidian-llm-wiki-local
  • Writer/viewer split — louiswang524/llm-knowledge-base
  • Multi-tool portable memory framing — remember-md/remember
  • Hosted-demo angle (to be used for Phase 6.5 Self-Demo, not the tool itself) — lucasastorian/llmwiki
  • Single-skill marketing — Astro-Han/karpathy-llm-wiki

Decisions informed by this research

  1. Keep llmwiki stdlib-first. lucasastorian/llmwiki shows the "full-stack" approach exists; llmwiki is the local alternative.
  2. Ship an Obsidian adapter in v0.1. Four of 15 reference implementations use Obsidian — clearly important to users. Make it an optional input adapter, not the only path.
  3. Ship the HTML viewer as the hero feature. None of the reference implementations have a beautiful static HTML output. This is llmwiki's most visible 10x.
  4. Keep the slash commands compatible with SamurAIGPT/llm-wiki-agent and kfchou/wiki-skills. Users can switch between implementations.
  5. Use Karpathy's three-layer structure exactly (raw/ immutable, wiki/ LLM-maintained, schema in CLAUDE.md/AGENTS.md). No deviations.
  6. Build-time redaction is non-negotiable — none of the reference implementations do this, and session transcripts leak PII by default.

References