New episodic memories are written continuously into three topic clusters. Each one holds an importance score that decays with time unless it is reinforced. Hybrid queries retrieve the nearest, most important memories, boosting them; memories retrieved often enough get pinned and stop decaying. When a cluster accumulates enough weak episodic memories, they are consolidated into a single semantic memory. If the store exceeds capacity, the least important unpinned memory is evicted.
imp(t+dt) = imp(t) - decay·dt (unless pinned)
score(q,m) = 0.6·sim(q,m) + 0.4·imp(m)
consolidate: weakest episodic subset -> 1 semantic node
evict: min(imp) among unpinned, when count > capacity
- Decay rate — how fast unreinforced memories lose importance.
- Memory capacity — the store's size limit before eviction kicks in.
- Query rate — how often hybrid retrieval queries arrive and reinforce hits.
- Consolidation interval — how often weak episodic clusters get summarized into semantic memory.
Sphere size and glow track importance; gold spheres are pinned, and a brief ring pulse marks each consolidation event.