1. System Architecture
Bio-mimetic pipeline for noise filtration and self-healing.
Novelty Filter
Vector comparison ($V_{new}$ vs $V_{db}$). Duplicates trigger Reinforcement Events instead of consuming new slots.
Sparse Tensor Memory
Fixed capacity ($N=1024$). The Eviction Gate continuously prunes low-utility vectors based on Confidence, Usage, and Decay.
Conflict Resolution
Semantic "Immune System". Contradictions trigger a Multi-Agent Debate (Advocate vs Challenger) to update truth confidence.
2. The Mathematics of Forgetting
The eviction score $E(m)$ is the heartbeat of SCTM. It mathematically mimics biological synaptic decay.
3. Memory Core Simulation
Observe the Eviction Gate operating in real-time.
Operations
4. Empirical Evidence
Benchmarking against Standard RAG & LRU Caches.
Evolving Facts Accuracy
Retrieval precision of changing facts over time.
SCTM Performance
Ranking
5. Ablation Studio
Toggle components to analyze their contribution.
Confidence Weighting
Protection of verified truths ($C$)
Usage Frequency
Reinforcement of active data ($log U$)
Temporal Decay
Ebbinghaus forgetting curve ($e^{-\lambda t}$)