Renvoo Memory System

This workspace now contains a from-scratch memory system designed for long-term use with an Obsidian-compatible vault for active reasoning and a Pinecone archive for large semantic recall. It is grounded in Karpathy's llm-wiki.md gist, vendored locally under vendor/karpathy-llm-wiki/.

What it fixes

The basic "Obsidian + LLM wiki" pattern in the transcript is useful, but it breaks down when the vault gets large. This build adds:

• Sharded indexes instead of one ever-growing global index.
• Semantic archive search with Pinecone for bulky or cold content.
• Drift and contradiction tracking through lint reports and source-linked notes.
• Monthly logs and compaction hooks so the working set stays small.
• A static agent profile separate from changing memory so identity does not bloat.

Layout

memory/
  .obsidian/               Obsidian-friendly vault settings
  assets/                  Downloaded images and attachments
  archive/                 Pinecone sync metadata and local cache
  raw/                     Original sources
  system/                  Stable identity, rules, and domain definitions
  wiki/
    contradictions.md
    log.md
    indexes/
    topics/
    entities/
    decisions/
    sources/
    maintenance/
    queries/

Quick start

1. cp .env.example .env
2. Add OPENAI_API_KEY if you want LLM-assisted synthesis and answers.
3. Add PINECONE_API_KEY if you want semantic archive search and long-term cold storage.
4. npm install
5. npm run memory:init
6. Open the memory/ folder as an Obsidian vault if you want the visual graph layer.

Daily workflow

• Ingest content:
  • npm run memory:ingest -- --file /absolute/path/to/file.md
  • npm run memory:ingest -- --url https://example.com/article
  • npm run memory:ingest -- --text "short note text" --title "Idea capture"
• Query memory:
  • npm run memory:query -- --question "What matters most about X?"
• Check health:
  • npm run memory:lint
• Rebuild indexes and compact the hot layer:
  • npm run memory:compact
• Inspect status:
  • npm run memory:status

Website

The repo now also includes a clinic-facing marketing site under src/site/.

• Start local website dev:
  • npm run site:dev
• Build the static website:
  • npm run site:build
• Preview the built website:
  • npm run site:preview

Website build output goes to dist/site/. If you set SITE_URL=https://your-domain.example when running npm run site:build, the build also generates sitemap.xml and a robots.txt file that includes the sitemap location.

Deployment notes live in docs/website/deployment.md.

Legal Pack

The repo now also includes a simple clinic-SaaS legal and compliance pack under docs/legal/.

• docs/legal/contracts/: pilot MSA, DPA, security annex, and pilot SOW templates
• docs/legal/privacy/: data inventory, retention matrix, processor register, breach SOP, DPIA-lite, and notice drafts
• docs/legal/corporate/: BV setup checklist and IP assignment templates
• docs/legal/regulatory/: non-clinical software classification memo

These are starter templates for Renvoo's Dutch dental-clinic pilot motion and should still be reviewed by Dutch counsel before signature or launch.

Content model

Keep only distilled, active reasoning in the wiki. Large transcripts, PDFs, exports, and archives belong in Pinecone-backed storage with summarized wiki notes pointing back to them.

That is the main architectural fix for the scaling problem in the original pattern:

• system/ is identity and stable rules.
• wiki/ is active reasoning and evolving judgment.
• archive/ is semantic recall for large and stable source material.

Notes

• If OPENAI_API_KEY is missing, ingestion falls back to a deterministic local draft flow.
• If PINECONE_API_KEY is missing, archive sync is skipped and query falls back to local vault search only.