v2.0.0 — capture → distill → recall

OpenExp v2

Complete architecture rework after the 8-month Q-learning experiment (removed — mean Q-value 0.006 across 27k memories; see README «Not a Q-learning memory system»).

Three layers

• Capture — byte-exact raw session archive (SessionEnd hook), permanent source of truth
• Distill — versioned, re-runnable: decision extraction + transcript ingest into Qdrant; nightly launchd pickup
• Recall — hybrid retrieval (semantic + BM25 + recency + importance), SessionStart top-N + per-message context hooks

Also

• Prediction → outcome journal (pure log, corpus for the experience-pack thesis)
• Daily/weekly/monthly retrospectives via launchd
• Experience packs: facts-only schema v3, anonymized trajectories, MCP prediction tools
• Pack catalog: CATALOG.md (5 published packs)
• MCP server: search_memory, add_memory, log_prediction, log_outcome, memory_stats