1.3 KiB
2026-02-25 Session Notes
Predictive Memory Scorer Architecture Review
Session opened with Nicholai reading and requesting comparison on the Predictive Memory Scorer architecture document. The document outlines a comprehensive system for training a per-user memory ranking model that evolves from heuristic baseline scoring (decay + BM25) to learned predictions using a lightweight cross-attention architecture.
Key architectural approach: standalone Rust binary sidecar communicating with daemon via JSON-RPC 2.0 stdin/stdout. Model is lightweight (~370K params, d=64 internal embedding dimension) trained on per-memory relevance labels from continuity scorer. RRF-based ranking fusion with earned influence (α) that adjusts based on session-end NDCG comparison wins. Cold start with pure baseline fallback, progressing to active prediction after 10+ sessions and >40% success rate threshold.
Plan includes 8 implementation phases: data pipeline prerequisites (session_memories table + per-memory scoring) → Rust autograd/model core → training pipeline → daemon integration → observability backend → dashboard UI → documentation. Full execution plan with delegation to 8 agents specified, with security audit and integration testing phases.
User attempted to share comparison image but attachment appeared corrupted.