Useful work with AI shouldn't disappear when a conversation ends. Wenlan builds the right pages and keeps them current as sources change, asking only when judgment is needed.
Get started
Desktop app
The desktop app is the fastest way to see the complete workflow: read pages, inspect their sources, and curate the knowledge system. The current macOS Apple Silicon preview is not yet notarized, so this installer verifies the GitHub release, installs Wenlan, clears quarantine for this app only, and opens it without changing macOS security settings:
/bin/bash -c "$(curl -fsSL https://raw.githubusercontent.com/7xuanlu/wenlan/main/scripts/install-macos-app.sh)"
The installer is inspectable. It checks the release archive against GitHub's published SHA-256 before replacing an existing app. Prefer the DMG or want to inspect the app source? See wenlan-app releases and wenlan-app.
Set up with your AI
Paste this into Claude Code, Codex, or another tool that can follow a setup guide:
Set up Wenlan for this AI client by following:
https://raw.githubusercontent.com/7xuanlu/wenlan/main/docs/setup-with-ai.md
Install only what this client needs. Then verify the local runtime,
its Wenlan connection, and a capture/recall round trip.
The guide detects which client you are using and keeps client-specific commands out of this README. It does not configure every AI tool unless you ask it to.
Need only the headless runtime on macOS Apple Silicon?
npx -y wenlan setup
This downloads the prebuilt CLI, daemon, and MCP connector, starts the local runtime, and verifies it. No Rust toolchain or Cargo is required. Linux x64/ARM64 with glibc has an automated shell setup path; Windows x64 uses the matching archive from Releases. macOS Intel currently has no supported complete-runtime install.
Manual and client-specific instructions: AI-assisted setup Β· Claude Code plugin Β· Codex plugin Β· CLI and MCP.
What is this?
Wenlan turns documents, notes, and past AI conversations into a source-backed knowledge base that stays current as your work evolves. Sources remain traceable; decisions, lessons, and corrections become durable memories; both can support the same maintained Pages.
Built for work that continues. Wenlan is for researchers, writers, consultants, product teams, and software teams whose knowledge is scattered across documents, notes, and AI conversations. It turns that material into inspectable Pages that can improve across projects and weeks, not another chat history or isolated memory store. It is not a life-management system or a memory SDK embedded inside another product.
One knowledge system, three roles:
- Sources keep the material Wenlan reads traceable. Imported conversations remain as captured records; registered files sync their current contents as they change.
- Memories preserve what work teaches you. Agents capture atomic decisions, lessons, corrections, and supersession with provenance.
- Pages compile current knowledge. Wenlan turns relevant Sources and Memories into source-cited Markdown you can reuse, refresh, and review.
The LLM-wiki foundation, extended:
- LLM-wiki v1: Karpathy defined immutable Sources, an AI-maintained Markdown Wiki, and a co-evolving Schema of rules for structuring and maintaining it. Wenlan implements that foundation with typed Memory fields and built-in rules for Page structure, provenance, citations, refresh, ownership, and review.
- LLM-wiki v2: Rohitg00 added a memory lifecycle. Wenlan makes that direction concrete with traceable Sources, agent-captured Zettelkasten-style atomic Memories (one complete idea each), and maintained Pages built from both.
Wenlan's distinctive move: Sources and atomic Memories independently support maintained Pages. Memory history preserves how knowledge changed; Page history shows which current evidence supports the synthesis. Machine-maintained Pages can rebuild from current support, while changes to human writing wait as reviewable revisions.
A knowledge graph that gets more useful over time
The entity-relation graph is one part of Wenlan's wider connected wiki. Knowledge Pages hold maintained synthesis, Entities anchor reusable people, projects, and concepts, Source Pages make imported or synchronized material inspectable, and atomic Memories preserve decisions and changes. They work through separate, explicit links: Page-to-Page wikilinks, Page evidence, Memory-to-Entity links, and directed Entity relations.
Within the entity graph, a configured enrichment model extracts typed Entities, observations, and directed relations from Memories. Entity linking and resolution reuse existing nodes instead of treating every mention as new; each Memory keeps its Source and can link to multiple Entities. How the connected model is stored ->
- Meaning and direction: Relations use a seeded vocabulary such as
uses, part_of, contradicts, and replaced_by; unknown types fall back to related_to and become reviewable vocabulary proposals.
- Strength and provenance: A relation can store confidence, an explanation, and its source Memory, so stronger and weaker claims remain distinguishable and inspectable.
- Communities that compound: Label propagation groups Entities by relation density, weighted by the relation count between each pair. These groups can organize optional corpus summaries while Entity links add retrieval context.
- Correction without erasure: Related claims, corrections, and explicit supersession stay inspectable together while original Sources and Memory history remain.
During retrieval, dense entity matching finds query-relevant entities. When eligible graph links exist, the default graph-memory stream boosts linked Memories as a third RRF signal. The path is data- and scope-dependent, and Space boundaries still apply. How the graph path works ->
Retrieval across words, meaning, and connections
Wenlan's core search is a local hybrid pipeline, not a single vector lookup. Each stage has a different job:
- Exact wording β SQLite FTS5: a full-text index finds literal terms, identifiers, and phrases.
- Similar meaning β FastEmbed +
Qdrant/bge-base-en-v1.5-onnx-Q: a quantized English model creates 768-dimensional embeddings; libSQL cosine DiskANN indexes them for approximate nearest-neighbor retrieval.
- Combined ranking β weighted RRF (
k = 60): lexical and semantic rank lists are fused without pretending their raw scores share a scale; cosine similarity also weights the vector contribution.
- Connected context β graph-memory stream: eligible entity links add a third RRF signal while the active read scope still filters returned Memories.
- Optional precision β cross-encoder reranking: unlike embeddings,
jinaai/jina-reranker-v1-turbo-en or BAAI/bge-reranker-base reads each query-candidate pair and reorders the smaller pool; reranking is off by default.
Page, episodic, and fact channels are opt-in and degrade to the remaining search signals if unavailable. Space still limits the read scope. Methods, defaults, and limitations ->
Two lifecycles, one maintained knowledge system
A generated wiki can go stale; a memory store can fragment into disconnected facts. Wenlan links two lifecycles without collapsing them into one layer.
Atomic Memory
CAPTURE -> CLASSIFY -> ENRICH -> LINK -> RECONCILE
Capture and explicit supersession are core. Model-backed stages run only when the matching model is configured, and the reconcile pass is off by default.
| Operation | What Wenlan does |
|---|
| Capture | Agents write one complete, self-contained idea per Memory, following the Zettelkasten atomic-note principle instead of saving the whole conversation. |
| Classify | With the on-device model, Wenlan assigns identity, preference, decision, lesson, gotcha, or fact; a precise type supplied by the caller remains authoritative. |
| Enrich | With the on-device model, adds structured fields, retrieval cues, event dates, quality, importance, and tags when available. |
| Link | Retains provenance and, when enrichment is enabled, connects Memories to entities and relations in the knowledge graph. |
| Reconcile | Explicit replacements preserve a supersedes chain. An optional on-device pass can queue protected conflicts for review instead of overwriting history; it is off by default and must be explicitly enabled. |
Advanced configuration: set WENLAN_ENABLE_DUAL_POOL_RESOLVE=1 to enable that reconcile pass.
Maintained Page
DISTILL -> CITE -> TRACK -> REFRESH -> REVIEW
| Operation | What Wenlan does |
|---|
| Distill | Compiles related Sources and Memories into one Markdown Page. |
| Cite | Retains citation records and verification status; automatic refresh discards a draft when its citation-support check fails. |
| Track | Records which evidence supports the Page, why it became stale, and a bounded changelog. |
| Refresh | When a Page is marked stale, rebuilds the eligible machine-maintained Page from current evidence. |
| Review | Turns changes to a Page you edited into a proposed revision instead of a silent rewrite. |
For example, import a design document and capture a debugging decision in Codex. Wenlan can compile one Page that cites both. When that Page is refreshed, it rebuilds from its current support; if you have edited it, the proposed change waits for review.
Local Markdown that works with Obsidian
Your durable synthesis remains ordinary files rather than a proprietary editor format:
- Plain files: Pages and session notes stay as Markdown under
~/.wenlan/.
- Inspectable history: Distill and handoff workflows can commit logical file batches to a local git repository.
- Obsidian coexistence: Wenlan reads an existing vault as a source. Symlink
~/.wenlan/pages/ into the vault or export a Page from the desktop app; your edits remain human-owned, and later machine refreshes become reviewable revisions.
The local history is directly inspectable:
$ git -C ~/.wenlan log --oneline
a1b2c3d distill: 4 pages
9f8e7d6 session: embedding-work
Capabilities
- Chat import: Bring in ChatGPT or Claude export ZIPs; Wenlan automatically skips conversations already imported.
- Document Sources: Ingest one
.md, .txt, or text-extractable .pdf file; recurse through a folder of them; or index Markdown from an Obsidian vault.
- Incremental sync: Regular file and folder Sources track changes in the background; Obsidian vaults stay read-only and resync on demand.
- Atomic Memory: MCP clients save one complete decision, lesson, correction, preference, or fact, with provenance and supersession recording where it came from and what it replaces.
- Typed enrichment: A configured model classifies each Memory, then adds the structured fields defined for its type, plus dates, tags, retrieval cues, and graph links.
- Source-backed Pages: Distill related Sources and Memories into Markdown Pages with source references and
[[wikilinks]]; the daemon can verify and record per-claim citations.
- Citation-gated refresh: Automatic refresh rejects citation-poor drafts; machine Pages update while human edits become reviewable revisions.
- Hybrid retrieval: FTS5 finds exact words, local BGE embeddings find meaning, and RRF fuses their ranks; graph links can add context.
- Retrieval channels: Optional Page, episodic, and per-fact channels widen recall; cross-encoder reranking can improve precision.
- Knowledge graph: Typed entities, relations, and observations connect people, projects, claims, and supporting Memories.
- Human-in-the-loop review: Routine work stays automatic; protected conflicts, Page revisions, entity merges, and new vocabulary wait for judgment.
- : Keep work, personal, client, and repository knowledge inside an explicit retrieval scope.
Daily workflow
The system above becomes a small daily loop: start with relevant knowledge, capture what matters while you work, close with a handoff, and let Wenlan refine what should return next time. Each pass leaves the same knowledge base sharper instead of creating another disconnected history.
The loop has four steps:
- Find current knowledge. Open a relevant Page, search, or use
/recall <query>; /brief [topic] reads the current Space Brief, and a topic appends separately labeled context from that same Space. Clients without plugin commands use the equivalent page, search, recall, and brief tools.
- Capture and find knowledge while you work.
/capture <thing> saves a decision, lesson, gotcha, or fact with its source. /recall <query> retrieves only what is relevant instead of loading your whole history.
- Close the loop.
/handoff records what changed and applies typed item-level updates to the current Space Brief.
- Keep the wiki current.
/distill deliberately creates or refreshes pages. Between sessions, optional model-backed passes can enrich captures, connect related entities, and refresh eligible pages. /lint checks knowledge health; /curate brings proposed revisions and any conflict-review items created by the optional reconcile pass to you.
Models and privacy
- Local base retrieval: The BGE embedding model runs through FastEmbed on your machine for hybrid search and needs no API key.
- Optional on-device synthesis: Enrichment and Page synthesis can use user-selected
Qwen3 4B or Qwen3.5 9B through llama.cpp. Wenlan does not download or activate a language model until you choose one.
- Other providers: An OpenAI-compatible local endpoint such as Ollama or LM Studio, or a configured cloud provider, can supply model-backed enrichment and synthesis instead.
- Cloud disclosure: If the model endpoint you select is remote, Wenlan sends that task's system and user prompts to it. Local retrieval and on-device synthesis stay on your machine.
- No telemetry: Wenlan sends no telemetry.
Full workflow reference: plugin/skills. Technical model roles: technical foundations.
Evaluation
This is a retrieval-only snapshot, not a claim about end-to-end answer quality. Method, environment receipts, and the update workflow live in docs/eval.
| Benchmark | Recall@5 | MRR | NDCG@10 |
|---|
| LME_Oracle (500 Q) | 93.6% | 0.857 | 0.883 |
| LME_S (deep, 90 Q) | 87.7% | 0.815 | 0.822 |
Learn more
More detailed documentation, concepts, and comparisons:
Docs
- Get started: install and verify the first local loop.
- Daily workflow: brief, capture, recall, handoff, distill, lint, and curate.
- MCP clients: connect Claude Code, Codex, Cursor, Claude Desktop, and other clients.
Concepts
Comparisons
Contributing
Bug fixes, eval cases, docs, and features are welcome. Installing Wenlan does not require building from source. For local development, run each group from the root of the named repository:
# 7xuanlu/wenlan β runtime, CLI, and MCP
cargo build --workspace
cargo test --workspace
# 7xuanlu/wenlan-app β desktop app
pnpm install
pnpm tauri dev
pnpm build:all
Use pnpm dev:all in the app repository when you want a fresh daemon-plus-app sequence. See this repository's AGENTS.md and CONTRIBUTING.md, plus wenlan-app's AGENTS.md, for the complete development workflow. Security reports: SECURITY.md. Please also read the Code of Conduct.
License
Wenlan is licensed under Apache-2.0. This includes the local runtime, CLI, MCP server, shared types, and Claude Code/Codex plugin files in this repository.
Lineage and peers
Wenlan (ζηΎ) takes its name from ζηΎι£, an imperial library that held εεΊ«ε
¨ζΈ as part of one of China's largest book collections.
Wenlan's llm-wiki v2 model is its own product direction, informed by the LLM-wiki and agent-memory lineages:
