From Retrieval to Resonance: Multi-Dimensional Episodic Memory Activation for Long-Horizon Human–AI Interaction

Six-String Resonance (SSR) reframes long-horizon AI memory as interactional memory activation: the memory that matters is not always the memory most semantically similar to the current query. SSR represents six continuity dimensions—semantic similarity, emotional congruence, narrative continuity, temporal rhythm, collaborative-state resonance, and unresolved commitment anchors—and evaluates them with SSR-Bench, a controlled diagnostic benchmark of 200 annotated episodic memories and 110 queries. Under gold interaction states, SSR reaches P@1 = 0.573 and perfect semantic-distractor resistance (P@1 = 1.000); removing the unresolved-anchor channel causes a 0.700 P@1 drop on anchor-targeted queries. Extended matched-baseline experiments locate a deeper source of signal in structured continuity representation: a metadata-aware LLM reranker reaches P@1 = 0.682, while fully inferred query states reduce SSR to P@1 = 0.345. Together, these results separate long-horizon memory into three coupled problems: continuity representation, state inference, and selective activation. Public preprint v1.2 includes extended evaluations of matched baselines, state inference, robustness to state noise, and adaptive weighting.