AI-native desktop learning app — a Tauri 2 shell that wraps your own local Claude Code CLI, grows a local-first markdown vault + knowledge graph out of your chat sessions, and drives concept review with FSRS-6. macOS, local-first, your data never leaves your disk.
Mneme combines, in one desktop app: an Obsidian-style local-first markdown vault (full data ownership), NotebookLM-style source-grounded chat with citations (anchored mode), Claude Code's full agent capabilities (MCP, web, tool use, hooks, skills) running as a subprocess inside a GUI shell, a mind-map layer for humans and a knowledge-graph layer for the AI, FSRS-6 spaced repetition on concept pages (not flashcards), and Echo360 lecture video in an embedded webview.
The end-experience it optimizes for: "this AI truly understands me" — proactively surfacing where you are, where you struggle, and how your knowledge connects, instead of only answering what you ask. The whole loop — learn → AI teaches → notes captured → reviewed via FSRS — feels like one product, not five glued together.
Compliance by construction: Mneme spawns your claude CLI under your subscription. No token proxying, no credential harvesting, no bundled model access.
Shipped: Tauri shell + streaming Claude chat (markdown/KaTeX/sanitized), local vault with hardened writer (path-canonicalization + symlink guards), manual file import, 6-step onboarding wizard, settings panel. Test baseline: 290 vitest + 73 cargo + svelte-check clean.
In progress next: multi-session sidebar, command palette, Tiptap editor. Full queue: BACKLOG.md; product identity: docs/PRODUCT.md.
Currently dogfood-stage software: built first for the author's own coursework, distribution to other users planned after it survives daily use. Expect sharp edges.
Requires: macOS, Node ≥ 20, Rust ≥ 1.88, and a working claude CLI on PATH (your own subscription).
gh repo clone r1ckyIn/mneme && cd mneme
npm ci
npm run tauri dev- Left course file tree · Middle video + material preview · Right Claude chat (streaming) · Bottom live course mind-map — final layout will be user-draggable (REQ-01 rework pending)
- Stack: Tauri 2 + SvelteKit (
adapter-static) +tauri-plugin-shell+ marked + KaTeX + DOMPurify + Svelte 5 runes;rusqlitevault index; vendored MITclaude-code-parser - Docs map:
docs/PRODUCT.md(identity) ·docs/specs/(capability specs) ·CONTEXT.md(glossary) ·docs/design/(visual SSOT) ·docs/adr/(decisions)
Mneme 在一个桌面应用里组合：Obsidian 风格本地优先 markdown vault（数据完全属于你）、NotebookLM 风格带引用的 source-grounded 对话（anchored 模式）、Claude Code 完整 agent 能力（MCP / 联网 / 工具 / hooks / skills）以子进程形式跑在 GUI 壳内、给人看的思维导图层 + 给 AI 看的知识图谱层、概念页（非卡片）上的 FSRS-6 间隔重复、以及嵌入 webview 的 Echo360 课程视频。
它优化的终局体验是："这个 AI 真的懂我"——主动浮出你在哪、卡在哪、知识怎么连，而不只是回答你问的。整条循环——学习 → AI 讲解 → 笔记自动沉淀 → FSRS 复习——是一个产品，不是五个 app 粘合。
合规即架构：Mneme 拉起的是你自己的 claude CLI、你自己的订阅。不代理 token、不碰凭证、不捆绑模型访问。
已交付：Tauri 壳 + Claude 流式对话（markdown/KaTeX/消毒渲染）、加固写入的本地 vault（路径规范化 + symlink 防护）、手动文件导入、6 步首次启动向导、设置面板。测试基线：vitest 290 + cargo 73 + svelte-check 零错误。
下一步：多会话侧栏、命令面板、Tiptap 编辑器。完整队列见 BACKLOG.md；产品身份见 docs/PRODUCT.md。
当前处于 dogfood 阶段：先为作者自己的课程学习而建，扛过日常真实使用后再向其他用户分发。会有毛刺。
需要：macOS、Node ≥ 20、Rust ≥ 1.88、PATH 上有可用的 claude CLI（你自己的订阅）。
gh repo clone r1ckyIn/mneme && cd mneme
npm ci
npm run tauri devCodename history: developed under codename
learn-osuntil 2026-05-07. · 项目曾以代号learn-os开发至 2026-05-07。