Wenlan
The second brain your agents write as they work: repo facts, client calls, training plans. Every page stays cited, linked, and current, so you don't have to.
First the hand-tended vault. Then agents bolted onto wikis: faster writing, same rot. Wenlan is what comes after.
Obsidian vaults, Notion pages, team wikis. Hand-tended, and alive exactly as long as your discipline.
Agents dump notes to Markdown files. No sources, no recall, no upkeep. It rots faster, with more confidence.
Agents write into your vault, but the gardening is still yours. Easier capture, same rot.
The llm-wiki that maintains itself: written during real work, cited to sources, found by hybrid recall, enriched while you sleep.
And while you sleep, the wiki gets better.
Wenlan turns repo facts, client constraints, source trails, and study notes into source-cited pages agents can brief, recall, and hand off.
Keep the codebase docs engineers actually update.
living engineering docs · 28 sources
Cutover blocks on the session-token backfill; run it before the flag flip.1
Legacy tokens stay valid for 30 days after cutover, then hard-expire.2
The next agent opens the same living docs an engineer would update.
Wenlan captures decisions as work happens, refines the wiki between sessions, and briefs the next session when it starts.
/capture · during the session
Work leaves a trail
/distill · between sessions
Repeats become a page
Agents also create pages directly when a topic deserves one right away.
/brief · next session
The next agent starts warm
Release runbook
cites /runbooks/release, /docs/architecture
The next conversation starts from the handoff instead of reconstructing the past.
Cleaned decisions and lessons become durable pages instead of buried chat logs.
Migration plan
6 sources cited
Runbook
8 sources cited
Architecture map
10 sources cited
People, projects, and pages come back linked to the memory and to each other, so recall arrives with its whole neighborhood.
Wenlan refuses unsourced pages. Every distilled claim cites the memory that produced it, so it can be traced back and re-checked.
# Architecture map
Search writes go through the ingest daemon; the API crate only reads.[1]
Low-confidence captures and contradictions surface for review instead of silently entering context. You step in only when it matters.
contradiction
low confidencenew: “staging deploys from main”
mem_0087: “staging deploys from release branches”
Between sessions, background passes link entities, grow matching pages, and fade stale memories.
+ linked Dana to checkout redesign
~ refreshed page: architecture map
− faded 3 stale memories
Tag captures and recalls with a space, so a client fact never bleeds into a side project brief.
space: work128
space: personal41
space: client-arden67
Memory, pages, and session artifacts live in real git history. Inspect, diff, revert.
a41f2c distill: architecture map
9d03b7 capture: rollback lesson
3c88e1 handoff: session close
Every MCP client reads the same shared memory through one local daemon. The files stay yours.
~/.wenlan
one shared memorylocal daemon · plain files
Captures stage in a fast local index, then solidify into Markdown pages you can open, diff, and keep.
Index · working memory for agents
Three ways in, one query. Any single route misses two of these.
staging
Markdown · lasting record for you
~/.wenlan/pages/architecture-map.md
# Architecture map
Writes go through the daemon. [^1]
[^1]: /docs/architecture
a1b2c3d page: architecture-map refreshed (4 sources)
Obsidian-compatible · grep-able · yours to move
×Files alone
Every recall re-reads whole folders. The context window pays for it.
×A database alone
Fast but opaque. Nothing you can open, diff, or move.
✓Index + files
One query for agents. Readable, versioned files for you.
Hybrid retrieval, measured
Hybrid retrieval finds the right local context without replaying chat history.
Run the harness yourself.Full chat replay
4,505 tokens / query
No retrieval, whole transcript in context
Wenlan retrieval
168 tokens / query
CE-reranked, 500-question snapshot
Retrieval-only snapshots. LME_S records 87.7% R@5 and 0.815 MRR on the stratified N=90 deep-S fixture. Token comparison is full replay vs retrieved context.
Wenlan is an LLM wiki for AI work. Agents capture what they learn, you add sources you trust, and the local daemon keeps source-backed wiki pages current across chats, tools, projects, and time.
Built-in memory stores what the AI decided was important. You usually cannot trace it, correct it, or use it from another tool. Wenlan keeps memory local, visible, correctable, and traceable. Readable pages, session logs, and project status are versioned in ~/.wenlan/.git/, and every distilled page cites the source memory IDs that produced it.
Hybrid retrieval combines vector search (BGE-Base-EN-v1.5-Q, 768-dim), FTS5, reciprocal-rank fusion, knowledge-graph context, and the local BGE reranker. LME_Oracle is 93.6% Recall@5, 0.857 MRR, and 0.883 NDCG@10 on the 500-question snapshot. LME_S is 87.7% Recall@5, 0.815 MRR, and 0.822 NDCG@10 on the stratified N=90 deep-S snapshot. The eval harness ships in the repo at crates/wenlan-core/src/eval/.
Yes. Wenlan runs on your machine and stores its database locally. No cloud sync or telemetry by default. Local memory setup works without a model or API key. On-device models or an Anthropic key are opt-in for automatic page distillation, recaps, and richer graph work.
No. The MCP server is the connector. Wenlan also includes the local daemon, manual /distill, optional model-backed background extraction and page work, a libSQL store with DiskANN vectors, FTS5 + knowledge graph, mandatory provenance, real git versioning for memory, page, and session artifacts, and readable Markdown export paths.
Claude Code and Codex have plugin paths. Cursor, Claude Desktop, VS Code, Gemini CLI, and other local clients connect through Wenlan's MCP server. ChatGPT and Claude.ai connect through Streamable HTTP MCP, with Remote Access in the desktop app providing the guided path. Remote Access has no authentication; anyone with the URL can access Wenlan, so stop Remote Access when unused.
No. Wenlan is not a notes app or a writing tool. It captures and refines what you learn from AI conversations. The Markdown projection under ~/.wenlan/ can be symlinked into Obsidian if you want to read it there.
Claude Code uses the marketplace plugin and /setup. Codex can run Wenlan through its plugin or through wenlan connect codex. Other local clients use wenlan connect <client>. ChatGPT and Claude.ai use the desktop app's Remote Access URL through Streamable HTTP MCP. The URL has no authentication, so treat it as a secret and stop Remote Access when unused.
Yes. The current prebuilt daemon release covers macOS Apple Silicon, Linux (x86_64, aarch64; glibc), and Windows (x86_64). macOS Intel has source/dev paths but no current prebuilt macOS Intel runtime. Service registration uses launchd on macOS, systemd-user on Linux, and Task Scheduler (schtasks) on Windows.
Yes. Memories, pages, and recalls belong to a space (for example, work, personal, or client-X). Set the active space per shell with WENLAN_SPACE, or declare them in ~/.wenlan/spaces.toml. The auto-detector also picks a space from the current repo or workspace.
Yes. Wenlan is open-source. The local runtime, CLI, MCP server, Claude Code plugin, and Codex plugin files in the Wenlan repo are Apache-2.0.
Open source
The local runtime, CLI, MCP server, Claude Code plugin, and Codex plugin are Apache-2.0.
Get release updates.