Dein persönliches
lokales Gedächtnis

# Project: Personal Local Memory (RAG) — "Second Brain"

You are setting up a fully local, private memory system on this computer.
Everything runs offline. No data leaves this machine.

## Your working rules
- Explain each major step in ONE plain-language sentence before doing it,
  so a non-programmer understands what is happening and why.
- Ask for my confirmation before: installing software, downloading models,
  or creating folders outside the project directory.
- If a tool is already installed (check first!), skip its installation.
- At the end, write a beginner-friendly README.md (in German) that explains
  how the system works and how to use it daily.

## Target architecture
1. **Storage:** SQLite database with the `sqlite-vec` extension for vector
   search and FTS5 for keyword search (hybrid retrieval).
2. **Embeddings:** Ollama running `nomic-embed-text` locally.
3. **Ingestion:** A Python script that watches `~/memory/inbox/` for new
   .md/.txt/.pdf files, chunks them (~500 tokens, 50 overlap), embeds them,
   and stores them in the database. Processed files move to
   `~/memory/archive/`.
4. **API:** FastAPI + Uvicorn on http://localhost:8600 with endpoints:
   - POST /ask     → question → hybrid search (top 5) → optional answer
                     generated by a local Ollama LLM if one is installed
   - POST /ingest  → save a text note directly into memory
   - POST /upload  → accept file uploads (.md/.txt/.pdf) and save them
                     into ~/memory/inbox/ for the ingestion watcher
   - GET  /        → serves the frontend
5. **Frontend:** ONE single self-contained index.html (vanilla JS, no build
   tools, no npm), with German labels and a friendly, minimal design.
   Contains:
   - a chat-style DIALOG area: the user asks in a text field, the answer
     appears as a conversation (generated by the local Ollama LLM if one
     is installed, otherwise the top matches), with the matched sources
     listed under each answer
   - a DRAG-AND-DROP zone (plus a classic file-picker button) that
     uploads .md/.txt/.pdf files via POST /upload into the inbox
   - a "Merk dir das" quick-save: small text field + one button that
     stores a note directly into memory
   - an EXPORT button that saves the current answer or the whole
     conversation as a clean, self-contained HTML document — print-ready,
     so the browser's print dialog turns it into a PDF; use clear
     headings so the same export works as a simple handout/presentation
   Comment the HTML/JS generously so the owner can later customize it.
6. **MCP server (bonus):** Register the retrieval function as a local MCP
   server so Claude Code / Claude Desktop can query the memory directly.

## Setup steps
1. Check the operating system (macOS or Windows) and adapt all commands.
2. Check for / install: Ollama, Python 3.11+ (prefer `uv` as the package
   manager). Ask before each install.
3. Pull `nomic-embed-text` via Ollama. Ask before downloading (~270 MB).
4. Create project folder `~/memory-system/` and data folders
   `~/memory/inbox/` and `~/memory/archive/`.
5. Write all code with generous comments in simple English.
6. Create a start script (`start.command` on macOS / `start.bat` on
   Windows) that launches Ollama check + ingestion watcher + web server
   with one double-click.
7. Test end-to-end: ingest a sample note, query it via the web UI, and
   show me the result.
8. Write the German README.md covering: what each component does, how to
   add memories (inbox folder OR web UI), how to search, how to start/stop,
   and where the data lives (single SQLite file — easy to back up).

## Important constraints
- 100% local and free. No cloud APIs, no accounts, no telemetry.
- Keep total code small and readable (~450 lines across all files).
- If anything fails, explain the error in plain language and propose a fix
  before acting.

Start now with step 1.