enquire-mcp is a long-term memory MCP server for AI agents, using the user's local Obsidian vault as its memory database. It turns the user's existing notes and documents into persistent, queryable agent memory; retrieval results retain their sources and expose freshness information, helping agents re-verify outdated facts while not locking context into a single model vendor. The product is MIT-licensed, vendor-neutral, read-only by default, and makes no outbound calls or sends no telemetry during operation.
In terms of use, enquire-mcp mainly serves: first, providing long-term memory for AI agents by connecting an Obsidian vault to any MCP-compatible client (such as Claude Code, Claude Desktop, Cursor, ChatGPT, Codex, OpenClaw), enabling agents to maintain semantic recall across sessions, models, and providers for meeting notes, journals, research logs, and decision documents; second, serving as a personal knowledge base/second brain, with hybrid retrieval supporting more than 50 languages, allowing one to ask a question in English about a Russian journal and get correct results; third, agentic RAG and context engineering, providing capabilities such as per-signal scores, HyDE query rewriting, and sub-question decomposition.
Main capabilities include: hybrid retrieval fusing BM25, TF-IDF, and multilingual embeddings (fused via RRF), with optional BGE cross-encoder reranking, and per-signal scores; the dense retrieval path scales using HNSW and int8 quantization. For document intelligence, Markdown and PDF/OCR enter the same hybrid retrieval path, PDF hits retain page citations ([page: N]), and OCR can recover scanned pages; Canvas is parsed as a typed graph; dedicated tools can execute a supported subset of Dataview LIST/TABLE and supported Obsidian Base filters. For freshness, each search hit can carry age_days and a stale flag, with an optional recency weight so agents prefer new knowledge. In addition, wikilink graph enhancement is provided, using in-degree to break ties only when candidate sets have the same RRF score; obsidian_get_communities provides GraphRAG-light-style topic community discovery. Built-in evaluation tools (NDCG / Recall / MRR) can measure retrieval quality on a self-built query set.
Usage: it can be installed globally via npm (npm i -g @oomkapwn/enquire-mcp), or use the npx -y @oomkapwn/enquire-mcp serve --vault <path> command to run it in any MCP-compatible client. The CLI provides the enquire-mcp configure command, which can generate accurate configuration snippets per vault and client (non-destructive, does not write content), for example --tier hybrid --client claude-desktop --vault <path>. The repository's examples/ directory provides configuration templates for Claude Desktop, Cursor, ChatGPT custom GPTs (remote MCP over HTTP), and other clients, as well as client-neutral playbooks for recall, evidence, freshness, synthesis, and safe write operations. There is also an MCPB Basic desktop package (provided in the v4.0.0-rc.7 GitHub Release), containing exactly 13 read-only tools, 0 prompts, and no write tools, monitor control, persistent/disk indexing, embedding models, PDF, or OCR functionality; it requires a compatible MCPB host with Node.js 22.13 or higher. Basic's recommended catch-all search lazily uses in-memory TF-IDF over live Markdown. Full hybrid retrieval is still available through npm/CLI installation.
Limitations and status notes: this branch records an unreleased source candidate version, and the candidate-specific npm commands and MCPB download URL are currently unavailable; for available artifacts, use the published GitHub releases or the published npm version. Stable installation commands and candidate notes are kept separate. The MCPB package's byte stream is not claimed to be reproducible, because the pinned upstream packager records metadata at packaging time; when exact bytes are needed, verify the published checksums. Real desktop GUI installation, signing, directory approval behavior, and directory listings remain maintainer-gated acceptance work. Regarding privacy, enquire does not make outbound calls or send telemetry during service, but the requested note content is still passed to the MCP client you connect to, after which cloud processing is governed by that client's own privacy terms. No additional information on fees, pricing, or specific platform support scope is provided in the given materials.
