Related work and positioning

Each ingredient exists elsewhere. What we built is the combination, designed around one score.

• Forward decay (Cormode et al., 2009) introduced the time-invariant key for decayed aggregates and sampling.
• Threshold Algorithm (Fagin, Lotem, Naor) and block-max WAND in search engines prune top-k scans with per-block upper bounds.
• Sorted sets (Redis) and priority queues give static-score heaps.
• Search engines (Elasticsearch function_score with decay) give recency-weighted ranking, recomputed per query.
• Time-series and stream processors give windows but not ranking or feedback.
• Causal inference (Granger causality, Bayesian networks, transfer entropy) studies attribution offline, over whole datasets, not as an online query.

Positioning

HeapDB is to "why did this happen?" what a vector database is to "what is similar?": a specialized store whose index is shaped around one kind of question.