RAG++ ChatGPT Memory Integration: Hard Specification

**Version**: 1.0.0 **Status**: LOCKED - Implementation Grade **Last Updated**: 2025-12-27 Build a conversation-memory substrate for Computational Choreography where ChatGPT export is ingested **without losing DAG truth**, exposed through retrieval that answers four question classes: 1. **Exact**: "What did I say?" (lexical, IDs, timestamps) 2. **Semantic**: "What else like this?" (dense vectors) 3. **Causal**: "How did we get here?" (DAG traversal) 4. **CC-Specific**: "What stabilized last time?" (motifs/invariants/decisions) ### Non-Goals (MVP) - No agent memory reasoning engine - No training or online embedding updates (except stats/annotations) - No perfect phase classification (rule-based first) The `mapping` in conversations.json contains nodes with parent/children links. Multiple children = regenerations/branching. If you flatten to single sequence, you destroy the most valuable property: alternative continuations and exact adjacency relations.