A Wikipedia-style, fully editable browser for the memory of your OpenClaw agents. Runs entirely on your own machine, reads and writes directly to
~/.openclaw/, never phones home.
OpenClaw's agents accumulate a lot of memory — daily dream diaries, identity
files, extracted chunks with embeddings, memory-palace pages synthesized from
session history. All of it lives as markdown and SQLite inside ~/.openclaw/,
but it's not designed to be read by a human.
This project renders that memory like an encyclopedia: article pages with summaries, a table of contents, wikilinks between articles, a "What links here" backlinks panel, search, categories, and — unlike a typical memory-viewer — a built-in editor so you can curate, correct, and expand the wiki by hand.
|
Article view — TOC, body with auto-wikilinks, "What links here" panel. |
Editor — split-pane markdown with live preview; safe writes leave a |
|
Setup & config — installed version, gateway health, models, plugins, skills (with bin/env checks), cron jobs, dependencies. |
Search — hybrid keyword + FTS5 search with snippet highlighting across every memory source. |
| Feature | Notes |
|---|---|
| Article pages | Wikipedia-style layout with TOC, summary, categories, "What links here" |
| Auto-wikilinks | Plain-text mentions of known article titles are linked automatically |
| Full-text search | FTS5 over memory chunks + title/body scoring, snippet highlighting |
| Live editor | Split-pane markdown editor with live preview, unsaved-changes warning |
| Create pages | New wiki pages land under agents/<agent>/memory/wiki/<slug>.md |
| Safe writes | Every save leaves a timestamped .bak-<ts> next to the file |
| Soft delete | Removed pages are moved to memory/.trash/, never unlinked |
| Multi-agent | Every agent's memory is indexed, filterable by namespace + agent |
| Filter / sort / group | URL-backed filters: sort A–Z, recent, agent, namespace; group by namespace/agent/letter; editable-only toggle |
| Setup & config viewer | One page showing installed version, gateway health, models, agents, channels, plugins, skills with requirement checks, cron jobs with last/next run, extensions, and dependency status |
| Live reload | chokidar watches the filesystem; the catalog rebuilds on change |
| Local-only | Binds 127.0.0.1; no auth because nothing leaves your loopback |
Memory is organized into Wikipedia-like namespaces so you can tell at a glance what kind of page you're looking at.
| Namespace | Source | Editable? |
|---|---|---|
Main |
agents/<id>/memory/wiki/*.md (Memory Palace) |
✅ edit, create, delete |
Identity |
agents/<id>/IDENTITY.md, SOUL.md, USER.md, etc. |
✅ edit only |
Diary |
agents/<id>/memory/*.md (dream diary) |
🔒 read-only (journal) |
Chunk |
memory/<agent>.sqlite → chunks table |
🔒 read-only (synthesized) |
Diary and Chunk are read-only by design: Diary entries are timestamped journal records, and chunks are regenerated by OpenClaw's embedding pipeline, so editing them would be erased on the next dreaming cycle.
Requires Node.js 18+ and a running OpenClaw install at ~/.openclaw.
git clone https://github.com/Noopy420/openclaw-wiki.git
cd openclaw-wiki
npm install
npm run build # builds the frontend once
npm start # serves http://127.0.0.1:4700Open http://127.0.0.1:4700 in your browser.
npm run devRuns the backend on :4700 and a Vite dev server on :4701 with hot reload.
Vite proxies /api/* to the backend.
All configuration is via environment variables. None are required.
| Variable | Default | Purpose |
|---|---|---|
OPENCLAW_HOME |
~/.openclaw |
Root of the OpenClaw data directory |
WIKI_PORT |
4700 |
HTTP port |
WIKI_HOST |
127.0.0.1 |
Bind address — keep on loopback, memory is private |
WIKI_CACHE_DIR |
~/.openclaw/.wiki-cache |
Reserved for future materialized cache |
Example (PowerShell):
$env:WIKI_PORT = 5000
$env:OPENCLAW_HOME = "D:\openclaw-data"
npm startFor the best experience, enable OpenClaw's memory-wiki plugin so the gateway
synthesizes curated wiki pages from the dream diary. In
~/.openclaw/openclaw.json:
"plugins": {
"entries": {
"memory-wiki": { "enabled": true }
},
"allow": ["memory-wiki", "..."]
}Then restart the gateway (openclaw doctor --fix or openclaw). OpenClaw
Wiki will pick up the new pages automatically via its file watcher — you don't
need to restart it.
On any editable article, click Edit in the top-right. You'll get a split-pane editor:
- Left pane: raw markdown
- Right pane: live preview, including wikilink resolution
Every save:
- Writes a timestamped backup (
<file>.md.bak-2026-04-23T04-36-06-216Z) - Overwrites the original file atomically
- Rebuilds the article catalog so the UI reflects the new state
Use Wikipedia-style double-bracket syntax inside your markdown. The left side is the target article id; the right side (optional) is the display label:
See also [[main.gordon.pizza-preferences|Gordon's pizza rules]].Plain-text mentions of known article titles are auto-linked at render time,
so you usually don't need to write [[…]] manually — it's there for when
you want to disambiguate.
Click Create new article in the sidebar. Pick the owning agent, give the
article a title, and optionally seed it with some initial markdown. The
file is written to agents/<agent>/memory/wiki/<slug>.md.
Articles in the Main namespace can be deleted from the editor. The file
is moved to memory/.trash/ (never actually unlinked), so recovering from
a misclick is just a mv away.
Navigate to Setup & config in the sidebar (or visit http://127.0.0.1:4700/setup) for a single-page dashboard showing:
- Installed OpenClaw version vs. the version that last touched the config
- Gateway health: is it reachable on its configured port, auth mode, bind address, control-UI origins, and whether device auth is disabled
- Models: every provider, their API key status (masked), every defined
model, and which one is the primary vs. fallback. Flags any referenced
model that isn't actually defined in
models.providers - Agents: the union of agents declared in config and agents present on disk, with their workspace paths
- Channels: Discord/Telegram/etc. — enabled, token status, guild count
- Plugins: whether each one will actually load (enabled ∧ in allowlist), and whether it has a config block
- Skills: scans
openclaw/skills/*/SKILL.md, parses the frontmatter, and reports for each skill whether required binaries are on PATH and whether required env vars / API keys are set. Filter by name - Cron jobs: each job's schedule, last run time, last status, next run time, duration, and delivery target
- Extensions: whether Python-based extensions have a built venv
- Dependencies:
better-sqlite3,node,python— installed or not, withhowToFixhints when something's missing
All read-only — it just reports what's on disk and what the gateway responds with. Nothing is written.
The home page has a filter bar with:
- Full-text filter — narrow by title/summary substring
- Sort — title A-Z, title Z-A, recently updated, oldest, agent, namespace
- Group by — flat list, namespace, agent, or first letter (Wikipedia-style alphabetical index)
- Agent — show only articles from one agent
- Namespace — show only Main / Diary / Identity / Chunk
- Editable only — hide synthesized and read-only pages
All filter state is stored in the URL, so bookmarks and links preserve your current view.
Everything the frontend does is available as a JSON API on :4700.
GET /api/status aggregate counts, agents, namespaces
GET /api/articles ?namespace= &agent= &category=
GET /api/articles/:id full article (markdown + rendered + sections + backlinks)
GET /api/articles/:id/raw raw markdown source
GET /api/articles/:id/backlinks { backlinks: [{id, title, namespace}] }
GET /api/search?q= hybrid keyword + FTS5 search
GET /api/categories [{name, count}] sorted by count
GET /api/recent most recently updated articles
GET /api/random { id } — random article
GET /api/agents [agent, ...] for the create form
GET /api/setup full config/state snapshot for the setup page
PUT /api/articles/:id body: {content} overwrite markdown
POST /api/articles body: {agent, title, content?} new page
DELETE /api/articles/:id move to .trash/
POST /api/reindex force a catalog rebuild
Write endpoints refuse to touch anything outside OPENCLAW_HOME and will
403 if the target article is in a read-only namespace.
server/
├── config.mjs env + paths
├── sources.mjs read markdown + SQLite chunks
├── articles.mjs normalize sources into Article model, build catalog
├── wikilinks.mjs entity index + [[wikilink]] injection
├── search.mjs hybrid keyword search
├── editor.mjs safe writes: save / create / delete
├── setup.mjs config + skills + cron + deps inspector (read-only)
├── routes.mjs REST endpoints
└── index.mjs Fastify server + chokidar watcher
web/
├── index.html
├── vite.config.js
└── src/
├── main.jsx router bootstrap
├── App.jsx routes table
├── api.js thin fetch wrapper
├── styles.css Wikipedia-Vector-inspired theme
├── components/
│ ├── Layout.jsx header, sidebar, footer
│ ├── SearchBar.jsx
│ ├── ArticleBody.jsx markdown renderer + wikilink resolution
│ └── TableOfContents.jsx
└── pages/
├── Home.jsx front page + filter/sort/group toolbar
├── Article.jsx read view
├── Edit.jsx split-pane editor
├── Create.jsx new-article form
├── Setup.jsx OpenClaw setup/config dashboard
├── Search.jsx
├── Category.jsx
├── Recent.jsx
├── Random.jsx
└── NotFound.jsx
The catalog lives in-memory and is rebuilt whenever a source file changes
(debounced 500ms). The rebuild traverses the agents directory, reads all
markdown + SQLite chunk groups, normalizes them into Article objects,
builds an entity index over article titles, and does a second pass to inject
wikilinks and compute backlinks.
For small-to-medium memory stores (a few thousand articles) this is
near-instant. For very large stores, the plan is to materialize the catalog
to wiki.sqlite — see the roadmap below.
- The server binds
127.0.0.1only. Do not expose it to a network — it renders and edits private agent memory with no access control. - SQLite is opened
readonly: true, fileMustExist: true. - Write endpoints refuse paths outside
OPENCLAW_HOME, refuse namespaces that aren't explicitly editable, and always write a backup before overwriting. - Deletes are soft (move to
.trash/).
The Chunk namespace requires better-sqlite3, which wraps the native
SQLite C library. It's declared as an optionalDependency so npm install
never fails because of it — if it's missing, you just lose that namespace.
The package ships prebuilt binaries for common Node versions:
- Node 18 / 20 / 22: works out of the box with
better-sqlite311.x. - Node 24: needs
better-sqlite312.x or later (12.x added prebuilt Windows binaries for Node 24). This repo targets^12.9.0.
If you hit a build error mentioning Visual Studio, it means npm fell back
to compiling from source. Either:
- Upgrade
better-sqlite3(npm install better-sqlite3@latest), or - Install Visual Studio Build Tools with the "Desktop development with C++" workload so the source build can complete.
The Setup & config page shows live dependency status — check there first if the Chunk namespace doesn't appear.
- Semantic search via
sqlite-vec(embeddings are already on disk) - Interactive knowledge graph view (Cytoscape/D3 force-directed)
- LLM-synthesized articles that cluster raw chunks into topical pages
when the
memory-wikiplugin hasn't curated any - History diff viewer using
chunks.updated_ator the.bak-*files - Per-agent color palette so multi-agent setups read at a glance
- Orphans / dead-links / wanted-pages "special pages" like Wikipedia
- Markdown editor upgrade to CodeMirror for syntax highlighting + folding
MIT — see LICENSE.




