Learning Loop

How it works

Capture

1. Parse. Claude's and ChatGPT's conversations.json exports, or a pasted transcript, are normalized into turns. For ChatGPT, only the branch you kept (walking up from current_node) is used — regenerated answers you abandoned are ignored.
2. 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.
3. 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.
4. Send. Highlights go to Readwise via POST /api/v2/highlights/ (or a CSV import) with source_type=learning_loop and .ai-chat tags.

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 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?