Cross-tool, cross-project session memory your coding agents can search and read — via MCP — while you keep a universal local vault of what those agents and tools already know.
Live demo · Karpathy's LLM Wiki pattern · MCP reference
llmwiki ingests sessions from Claude Code, Codex, Cursor, OpenClaw and others — plus files, folders, URLs, and PDFs you feed in — into a redacted markdown vault. Agents use that vault through the stdio MCP server (python3 -m llmwiki.mcp): search, read pages, health/lint, sync, export, and add sources without shelling out. Memory spans tools and projects without living inside a single chat UI. Put the vault somewhere you trust — it is the secure place for that unified store. Humans use the static site for the same vault: open site/index.html after build to check pipeline state, token and cache metrics, health, usefulness, and which entities and concepts have settled so later sessions can take them as given. Nothing has to stay running for MCP or for reading those files.
- MCP tools for agents —
wiki_search,wiki_read_page,wiki_health,wiki_sync,wiki_export,wiki_addover stdio (JSON-RPC). Wire it once in Claude Code, Cursor, Codex, or any MCP client. See docs/reference/mcp.md. - Scriptable source intake —
llmwiki add(CLI) orwiki_add(MCP) turns files, folders, URLs, and PDFs into markdown structured for free LLM digest, so bookmarks, cron jobs, or other agents can feed the same vault automatically. Once synthesised, that material is ordinary wiki content — same layer as session-derived pages. - A knowledge layer —
synthwrites source summaries and harvests entity/concept candidates. You review those (site/candidates.htmlor CLI) and apply decisions withllmwiki candidates apply. That review is a real gate, not an optional extra. - A static site for humans — Home, Analytics, Candidates, Topics, graph, and
Cmd+Ksearch split into Wiki and Site results. See what the pipeline did, browse settled entities/concepts, and search wiki text with the same literal-match rules agents use. Generated byllmwiki build; see the UI reference. - A vault you place — choose a secure directory for the unified store (raw sources, wiki pages, built site). It stays on your machine under a path you control; it is not mixed into the package install.
Page kinds: docs/reference/page-kinds.md. MCP tools: docs/reference/mcp.md. Site screens: docs/reference/ui.md.
Requires Python ≥ 3.12 (CI verifies 3.12 only; newer interpreters are installable but not matrix-tested). Runtime dependency is markdown only.
pip install llm-wiki-plus
llmwiki init --vault ~/llmwiki-vaultPoint a gitignored config.json at the vault so later commands need no --vault:
{
"vault": { "default_path": "/home/USER/llmwiki-vault" }
}From a clone, ./setup.sh installs the package and reports which agents it can see. The clone itself is not a vault — bare llmwiki init / sync / build against this repository are refused; name --vault or demo.
Agent kit (slash commands and skills) after install:
llmwiki install-agent-kit --dest ~/.claudeThat is what a pip user runs so /wiki-sync and friends work from their project. Point your MCP client at python3 -m llmwiki.mcp (same package). See docs/reference/cli.md and docs/reference/mcp.md.
The install also drops a small .llmwiki-agent-kit.json manifest in --dest, recording which files it wrote and their checksums. It is safe — and useful — to commit alongside .claude/commands/: it is what lets a later upgrade recognise its own files and remove ones it has retired, while leaving anything you wrote or edited alone.
Keep the vault current so agents (MCP) and humans (site) see the same knowledge:
sync / add → synth → review candidates → build
llmwiki sync # new transcripts → vault/raw/sessions/
llmwiki add notes.md # markdown file → vault/raw/docs/
llmwiki add https://example.com/post # web page
llmwiki add ./paper.pdf # PDF
llmwiki add ./research-folder/ # folder of docs
llmwiki synth # source pages + harvest candidates
# review candidates (CLI apply, or open vault/site/candidates.html after a build)
llmwiki build # vault/wiki/ → vault/site/ for human browse + metricssynth does not rebuild the site. build does — run it when you want Home, Analytics, and HTML refreshed. candidates apply rebuilds after a successful batch so /candidates.html matches the wiki (--no-rebuild skips that). Candidate review is the human gate: a row starts with no decision, Apply assembles llmwiki candidates apply --vault … --actions '…', and only that command writes the wiki.
One-shot: llmwiki all --graph-engine builtin runs sync → synth → build → graph → lint. Opt out of a stage with --no-sync, --no-synth, --skip-graph, or --skip-lint.
Look around when something feels off: llmwiki doctor diagnoses the install and vault from a bundled heal catalog and, with consent, runs safe repairs (--check is diagnose-only; see CLI reference). llmwiki lint stays the wiki-quality rule runner.
One row per agent. Supplies sessions = llmwiki sync can read its transcript store. Reads the wiki = MCP and/or slash commands after install-agent-kit (any MCP client can call python3 -m llmwiki.mcp). Core adapters run on a default sync when their store exists; contrib adapters need --adapter <name> or adapters.<name>.enabled in config.
| Agent | Supplies sessions | Reads the wiki | Core or contrib |
|---|---|---|---|
| Claude Code | ~/.claude/projects/*/*.jsonl |
slash commands, MCP | core (claude_code) |
| Codex CLI | ~/.codex/sessions/ |
AGENTS.md workflow, MCP |
core (codex_cli) |
| ChatGPT | contrib store | MCP | contrib (chatgpt) |
| Copilot Chat | VS Code workspaceStorage | MCP | contrib (copilot_chat) |
| Copilot CLI | ~/.copilot/session-state/ |
MCP | contrib (copilot_cli) |
| Cursor IDE | workspaceStorage | MCP | contrib (cursor) |
| Cursor CLI | ~/.cursor/chats/ |
MCP | contrib (cursor_cli) |
| Gemini CLI | ~/.gemini/ |
MCP | contrib (gemini_cli) |
| Obsidian | markdown vault intake | wikilinks in the vault | contrib (obsidian) |
| OpenCode | shared OpenClaw schema | MCP | contrib (opencode) |
| OpenClaw | ~/.openclaw/agents/*/sessions/ |
MCP | contrib (openclaw) |
llmwiki adapters --wide
llmwiki sync --adapter openclawShipped defaults: examples/sessions_config.json. Personal overrides: gitignored config.json at the clone root, merged on top. Full keys: docs/configuration.md.
Skip projects or dates with a .llmwikiignore next to the vault (one path or glob per line).
| Topic | Link |
|---|---|
| Hub | docs/index.md |
| Install + first vault | docs/getting-started.md |
| MCP server | docs/reference/mcp.md |
| Product principles | docs/maintainers/principles.md |
| Page kinds | docs/reference/page-kinds.md |
| CLI | docs/reference/cli.md |
| Site UI | docs/reference/ui.md |
| Architecture | docs/architecture.md |
| Upgrading | docs/UPGRADING.md |
| GitHub Pages | docs/deploy/github-pages.md |
This repository's CLAUDE.md and AGENTS.md are for people changing llmwiki itself. Vault workflows for users live in the installed agent kit.
- Andrej Karpathy — LLM Wiki idea
- Pratiyush — original llm-wiki project this repository builds on
MIT © Alexander Makarov; originally based on Pratiyush/llm-wiki