Persistent AI agents — systems that maintain identity, memory, and behavioral continuity across clean session boundaries — face a fundamental architectural tension: they either carry all past context (token-bloated and expensive) or they possess no automatic recall outside explicit retrieval calls (fragmented and forgetful). Human memory solves this through associative retrieval: relevant past experiences surface naturally when triggered by current context, while irrelevant traces remain dormant. We present the Tidal Layer, a lightweight associative memory architecture that bridges episodic and semantic storage through vectorized conversation embeddings tagged with emotional valence metadata. Every user-agent exchange is embedded and stored with its Emotional Valence Vector (EVV) state at time of creation. The architecture centers on a **Unified Knowledge Index (UKI) ** — a combined FTS5 keyword index and 384-dimensional vector store spanning agent skills, knowledge base documents, and tidal conversation memories. A pre-LLM module (`warmₘemory. py`) runs parallel FTS5 and vector similarity queries on every user turn, retrieving relevant context without explicit model invocation. The architecture includes exponential decay weighting, a dual-vector scorer that blends semantic similarity with emotional valence proximity, and adaptive weighting that shifts retrieval priority toward emotional resonance when affective intensity exceeds a threshold. A production implementation integrated within the broader agent architecture demonstrates the system running in continuous operation over 30+ days.
Robert Joseph Vandelinder (Mon,) studied this question.