Learn from sessions

moradin:learn-from-sessions

Mine your accumulated Claude Code session history for implicit preferences. Surfaces candidates for review; operator decides what becomes memory.

When to invoke

/moradin:learn-from-sessions [--since 180d] [--project <name>]

Best run ONCE at first-time setup, then again periodically (monthly or quarterly) to catch new patterns. For per-session incremental capture, use /moradin:ship after each session.

What you do

1. Set up. Confirm ANTHROPIC_API_KEY is in env. If not, ask operator to set it or use --local-only mode.

2. Discover sessions. Run scripts/session_scanner.py --since 180 --count-only to show operator how many sessions/turns will be scanned.

3. Estimate cost. Roughly $0.001 per turn for pre-filter + $0.01 per extracted candidate. For ~10k turns: typically $8-15.

4. Confirm with operator before running the expensive extraction.

5. Run extraction.

   python scripts/extract_preferences.py --since 180 --min-count 5 --min-sessions 3
   

   This:

   • Scans matching JSONL files
   • Pre-filters with regex (corrections / preferences / lessons signal patterns)
   • Classifies filtered turns with Haiku (cheap)
   • Extracts structured candidates with Sonnet (strong)
   • Clusters similar candidates
   • Applies stability thresholds (count ≥ 5, sessions ≥ 3, confidence ≥ 0.6)
   • Writes review file to scratch/proposed_<date>.md

6. Surface results to operator. Read the proposed file, summarize:

   • Total candidates that passed stability
   • Top 5 by occurrence count
   • Distribution: how many principle / preference / lesson candidates

7. Walk operator through review (optional). For each candidate ask:

   • ACCEPT → write to memory/preferences/<title>.md (or principles/ or lessons/ based on category)
   • EDIT → rewrite statement, then accept
   • REJECT → skip

8. After approval session. Run python scripts/refresh_indexes.py to update _INDEX.md files. Delete the review file.

Tuning knobs

Flag	Default	When to adjust
--since N	180 days	Start with 180 to verify pipeline quality; expand to 365+ once trusted
--min-count N	5	Lower (3) if your corpus is small; raise (8-10) if you want only strongest signals
--min-sessions N	3	Same logic — diversity threshold
--project <slug>	(all)	Restrict to one project (e.g. C--Users-yuezh-Documents-GodTech)
--dry-run	off	Test the regex pre-filter without spending LLM dollars

Privacy

• All session reading is LOCAL (Python reads files on disk).
• LLM calls go to Anthropic API. Per Anthropic's commercial terms, API traffic is NOT used for training.
• If you need pure-local processing: set MORADIN_LOCAL_ONLY=1 (requires Ollama installed; future enhancement).
• The script applies regex redaction for common secrets (API keys, emails, bearer tokens) before sending to LLM.

Don't

• Don't run this without operator confirmation on the LLM spend.
• Don't auto-write extracted candidates to memory/. Operator reviews each.
• Don't include any candidate in memory/ without first cleaning the title + statement.
• Don't run on a corpus you haven't redacted appropriately if it contains sensitive IP.

Output

• scratch/proposed_<date>.md — reviewable candidates
• 0 or more new files in memory/preferences/, memory/principles/, or memory/lessons/ after operator approval
• Updated _INDEX.md files