Wenlan Builds a Local, Source-Backed Knowledge Base for AI Agents

Summary

Wenlan is an open-source knowledge-base project for work that continues across AI sessions. It runs as a local daemon, bundled into a macOS or Windows desktop app or available through a headless CLI, and connects Claude Code, Codex, Cursor, Claude Desktop, and other clients through plugins or MCP. The system separates traceable Sources, atomic Memories captured by agents, and maintained Markdown Pages that synthesize both with citations. Its local retrieval pipeline combines SQLite FTS5 for exact terms, 768-dimensional FastEmbed embeddings indexed with Qdrant or libSQL DiskANN, and weighted reciprocal-rank fusion; optional graph links and cross-encoder reranking add connected context or precision. Memories can be classified and enriched with a configured on-device or remote model, linked to entities, and reconciled through explicit supersession and reviewable conflicts. Pages can be distilled, citation-checked, refreshed from current evidence, and staged for human review when edits affect human-owned writing. Data remains inspectable in local Markdown and libSQL storage, with Obsidian support, local health checks, and an offline write outbox. Wenlan says base retrieval is local and does not activate a language model until the user chooses one; remote model endpoints receive the prompts for their configured tasks. Its retrieval-only evaluation reports Recall@5 of 93.6% and MRR of 0.857 on one 500-question benchmark, while the project notes that this is not an end-to-end answer-quality claim.