Pensieve.

A practical guide to building a personal research automation system — from universal inbox to automated daily digest.

April 2026
22,000+
Book Highlights
784
Markdown Files Indexed
3
Search Systems
How it works

What This Looks Like in Practice.

A single query, three search systems in parallel, one synthesized answer under a minute.

"Find me all the historical analogies I've read about the current closing of the Strait of Hormuz"
Obsidian Vaults · Local Markdown Files
Qdrant
Semantic search

Searches 22,000+ book highlights by meaning. Finds passages about the Suez Crisis, the Tanker War of the 1980s, British naval blockades, oil embargoes — without the words "Strait of Hormuz" ever appearing.

Filters by genre, tag, or author. Sub-2-second search.

QMD
Hybrid search

Searches 784 markdown files across articles, notes, and newsletters. Runs three strategies simultaneously: keyword (BM25), semantic, and HyDE — hypothetical document expansion that imagines what a relevant passage would look like.

Surfaces articles, podcast transcripts, and personal notes.

Readwise
Full library

Searches across everything — including books and articles not yet indexed locally. Catches recent saves you haven't highlighted yet, or sources outside indexing filters.

The safety net for completeness.

The point: This isn't keyword search. I never tagged those highlights with "Hormuz." I never filed them under "geopolitics." The system finds connections across your entire reading history because it understands meaning, not just words.
Architecture

Three Systems, One Orchestrator.

You don't need all of them. Start with System 1 and add complexity only when you need it.

System 1 — Essential
Readwise Reader
~$12/month

A read-it-later service that stores everything you read, lets you highlight it, and makes it all searchable via API. The memory layer.

  • Email forwarding address — forward any email to become a highlightable document
  • Browser extension — save paywalled Substack articles while logged in
  • PDF upload — drag PDFs into the web app or email them in
  • Highlight sync — Kindle, Apple Books, web, PDFs, all unified
  • API access — Claude searches programmatically
System 2 — Advanced
Local RAG (Qdrant + Voyage AI)
Free, runs locally

A local vector database for semantic search across 22,000+ highlights. Probably overkill for most people — the most complicated piece.

  • Readwise syncs highlights to local markdown (via Obsidian)
  • Voyage AI converts each highlight to a 1,024-dimension meaning vector
  • Qdrant stores vectors locally — search by meaning in under 2 seconds
  • Metadata filters: genre, author, favorites, tags
System 3 — Optional
QMD
Free, local GPU

Local search engine for markdown files with three modes: keyword (BM25), semantic (vector), and HyDE. Runs on your Mac's Apple Silicon GPU.

  • Indexes articles, podcast transcripts, newsletters, and personal notes
  • Hybrid search combines all three strategies simultaneously
  • Complementary to Qdrant: QMD has everything that isn't books
Orchestrator
Pensieve
Claude Code skill

A reusable prompt + workflow that takes something you've read and produces a polished synthesis review.

1
Pull & Extract
Pull highlights, extract key claims, group by theme
2
Search & Connect
Search entire library for supporting, contradicting, and related work
3
Stress-Test
Is the throughline clear? Are connections earned?
4
Draft & Edit
Write synthesis → structure pass → humanizer → copy-edit
5
Finalize
Polished review essay with citations back to specific sources
Each phase writes its output to a file before the next phase starts. You can pause, review, redirect, or resume at any point. If Phase 2's search is good but you want to rewrite Phase 3, the research is already saved to disk.
Getting started

Three Tiers.

Start with Tier 1 in an afternoon. Add tiers as you need them.

Tier 1
The Foundation
1 afternoon

Get all your research into one searchable place.

  1. Sign up for Readwise Reader
  2. Get your email forwarding address
  3. Set up auto-forwarding rules for your sources
  4. Install the browser extension
  5. Get your Readwise API token
Monthly cost ~$12
Tier 2
Claude-Powered Summaries
1 evening

Ask Claude to summarize and synthesize your research on demand.

  1. Install Readwise MCP for Claude Code — Claude searches your library directly
  2. Or install the Readwise CLI for command-line access

Then you can say: "Summarize everything I've read about tariffs this month."

Additional cost $0
Tier 3
Automated Daily Digest
The real goal

Every morning, Claude automatically produces a digest of your recent research and emails it to you.

  1. Set up Readwise MCP connector on claude.ai
  2. Set up Gmail MCP connector
  3. Create a scheduled agent with /schedule
  4. Define your digest prompt (template below)

Runs in Anthropic's cloud. Fires on a cron schedule. No machine needs to be on.

Additional cost $0
Optional Tier 4 — Local RAG for deep research. After 2-3 months with a meaningful corpus, if you want to cross-reference everything ("what has every analyst said about India's manufacturing sector over 6 months, and where do they disagree"), add Qdrant + Voyage AI. The value scales with the size of your library.
Reference

Tools & Costs.

Everything you need, linked and priced.

Tool Purpose Link Cost
Readwise Reader Universal inbox for all reading readwise.io/read ~$12/mo
Readwise API Token Lets Claude access your library readwise.io/access_token Included
Claude Code AI assistant with MCP + scheduling claude.ai/code Subscription
MCP Connectors Connect Readwise, Gmail, etc. claude.ai/settings/connectors Free
Voyage AI Embeddings for RAG (Tier 4) voyageai.com Free tier
Qdrant Local vector database (Tier 4) qdrant.tech Free
QMD Local markdown search engine github.com/tobi/qmd Free
qpdf Decrypt password-protected PDFs brew install qpdf Free
No Readwise?

The Folder-Based Alternative.

Own the full pipeline with just Claude Code and local files.

~/Research/
├── inbox/           ← drop new PDFs, saved articles here
├── processed/       ← Claude moves files here after reading
├── highlights/      ← extracted key passages (markdown)
└── digests/         ← generated summaries
Readwise Path Folder Path
Setup effort 30 minutes 2–3 hours
Ongoing effort Almost zero (auto-forwarding) Some manual file management
Scheduled digests Works natively (MCP) Needs GitHub sync workaround
Cross-device Yes (phone, tablet, laptop) Local machine only
Cost ~$12/month Free
Data ownership Readwise holds your data Everything local
Bottom line
For most people, Readwise is the right call — the time savings on ingestion alone justify the cost. But if you want to own the full stack and don't mind upfront setup, the folder approach works.
The big picture

Architecture Summary.

Your Sources
Books & Kindle
Highlights, annotations
Articles & Newsletters
Substack, RSS, blogs
PDFs & Reports
Research, whitepapers
Podcasts & Video
Transcripts, notes
Notes
Meetings, personal, clients
→
Ingestion
Email forwarding
Auto-forward rules
Browser extension
Save while logged in
Manual upload
PDFs, decrypt if needed
YouTube URL
Transcript extraction
→
Storage
Readwise Reader
Universal inbox
Obsidian Vaults
Local .md files
→
Synthesis
Claude Code
On-demand or scheduled
Daily Digest
Weekly Cheat Sheet
On-demand “brief me”
The key insight: Obsidian is the single funnel. Everything ends up as local markdown files. Claude searches across those vaults by meaning. You never need to manually re-find or re-read anything — your entire corpus is always at hand.