Your conversations
Load the sample, upload a conversations.json export (Claude or ChatGPT), or paste a transcript. Nothing leaves your browser unless you choose to send it.
Highlights to keep
0Send straight to my Readwise
If your browser blocks the request, use the CLI in the repo: node cli/learning-loop.mjs export.json --push
What are you working on this week?
Daily Review today asks one question: which highlights are due? This adds a second: which due highlights matter to what I'm doing now?
Today: spaced repetition only
With this week's context
Your highlights, choosing your next book
Most bookstores open on a bestseller list. Readwise Books can open on something no other store has: what you've already highlighted, plus what you've been learning in AI conversations. Every pick below says which of your highlights it came from.
Store default: bestsellers
For you, from your highlights
How it works
Capture
- Parse. Claude's and ChatGPT's
conversations.jsonexports, or a pasted transcript, are normalized into turns. For ChatGPT, only the branch you kept (walking up fromcurrent_node) is used — regenerated answers you abandoned are ignored. - Score sentences. Each sentence of each answer gets transparent signals: definitions, contrasts ("not X but Y"), rules of thumb, causal claims, bold emphasis, whether it answered your question, and — the strongest signal — whether your next message reacted ("oh interesting", "didn't know that", "原来如此"). Filler ("Great question!") and sentences that depend on context ("It matters most when…") are dropped.
- Pick a few. At most two per answer and four per conversation, with near-duplicates suppressed. Your question becomes the highlight's note, so the review card has context.
- Send. Highlights go to Readwise via
POST /api/v2/highlights/(or a CSV import) withsource_type=learning_loopand.ai-chattags.
Context-aware review
final = w · relevance(context) + (1 − w) · dueness, where relevance is BM25 over highlight text, title and tags, and dueness is days since last review ÷ current interval (capped at 1). The list is then re-ranked with Maximal Marginal Relevance so one book can't take over the session. Chinese text is tokenized with character bigrams, so it works without a segmenter.
This is the same shape as an ad or feed ranking stack — a relevance model, a business constraint (spacing), and a diversity pass — scaled down to one reader.
What to read next (Readwise Books)
final = wtaste · taste + wintent · intent + wgap · gap, then MMR over topics.
- Taste: BM25 affinity between each of your highlights and each book, averaged over the three strongest highlights. Those three are kept as the explanation (“because you highlighted…”).
- Intent: how well the book fits what you typed for this week, in English or Chinese.
- Gap: the share of that evidence that came from AI conversations. If most of what you know about auctions came from Claude, a real book on auctions is the natural next step.
- Already read: books you already have highlights from (matched on title or translated title, plus author) are removed instead of being sold to you again.
This is the commerce version of the same ranking shape: a home feed that blends long-term affinity with session intent, keeps one category from taking over, and says why every item is there.
What a production version would change
- Embeddings instead of BM25 for relevance (Readwise already built hybrid search for MCP).
- Context inferred, not typed: what's in your Reader inbox this week, your recent tags, your calendar.
- An evaluation loop: does context-aware review raise "Keep" / "Mastery" rates and 30-day retention versus control?
- For Books: rank the real catalog with embeddings, add reading-session signals (finished, abandoned, re-read), and test the "for you" shelf against bestsellers on sample opens, purchases and books finished.