19 lines
1.5 KiB
Markdown
19 lines
1.5 KiB
Markdown
# 2026-02-25 Session Notes
|
|
|
|
## Predictive Memory Scorer Evaluation Session
|
|
|
|
Nicholai initiated a session to evaluate the predictive memory scorer design document (`docs/wip/predictive-memory-scorer.md`) against the X/Twitter algorithm reference code in `references/x-algorithm/`.
|
|
|
|
The session began with reading the 1,775-line predictive scorer design doc, which outlines a comprehensive architecture for training a per-user learned memory ranking model using:
|
|
- Cross-attention architecture (inspired by ACAN, Memory-R1, and X's Phoenix ranker)
|
|
- ListNet-style listwise ranking loss with KL divergence
|
|
- HashTrick tokenizer (16K buckets, 64-dim internal embedding space)
|
|
- ~1.11M total parameters (sub-2M target)
|
|
- Reciprocal Rank Fusion (RRF) for blending baseline and predictor scores
|
|
- Cold start handling with gradual influence ramp (sessions 1-10: 0.2 cap, 11-20: 0.4 cap)
|
|
- Topic diversity decay (0.5 decay factor for >0.85 cosine similarity overlaps)
|
|
- Behavioral training signals from FTS hit overlap during sessions
|
|
- Drift detection via success rate EMA and replay buffer for recovery
|
|
- Comprehensive observability (diagnostics domain, latency tracking, error ring, timeline events)
|
|
|
|
Two Explore agents were spawned to examine the X algorithm reference code in depth (architecture, scoring mechanisms, candidate pipeline, training signals, cold start patterns, feature engineering, and the Phoenix/Thunder/HomeMixer components). The exploration work was initiated but remains in progress. |