Salience Decay

Salience falls out of decay — no parallel salience substrate needed. Lyswen performs a ballad each day; between days, every belief weight is multiplied by a fixed rate. The ledger develops shape: recent performances near full weight, ancient ones faded. Identity beliefs tagged self_core are exempt and stay pinned at 1.0. Reinforcement is implicit — re-performing the same ballad creates a fresh high-weight edge alongside the decayed old one; the aggregate recovers without an affirm step.

Controls weight × rate

rate 0.90
Ledger (newest first — click a row to re-perform that ballad)
No performances yet.
Self-core (exempt from decay)
(lyswen, is, bard)self_core
1.000
Aggregate (all): 0.000
Aggregate (last 3 days): 0.000
Day: 0
Entries: 0
How it works. Each “Skip a day” applies apply_decay(graph, rate)edge.weight *= rate for every non-exempt agent edge. With rate=0.90 an edge halves in ~7 days; with rate=0.99 it halves in ~70. Edges flagged self_core skip the sweep entirely — identity anchors don’t fade. Re-committing the same SPO creates a fresh edge alongside the decayed one; an aggregator that sums weights gets a recency-biased read of “what’s happened lately” for free.

Tie to Eddy. This is memory/decay.py:apply_decay over a per-agent belief graph — exactly what examples/cognition_salience_decay.py runs. The recency curve isn’t a salience system; it’s the geometry of multiplicative forgetting. The next lab (Trust Modulation) shows the same weight × rate machinery in reverse — a positive multiplier on a single edge each time advice pays off.