A Dual-Path Associative Memory System with Topic-Aware Compression for Conversational AI

We present a dual-path associative memory system designed to enhance context understanding in conversational AI when user intent is ambiguous. The system employs two independent retrieval paths operating in parallel: a dense vector similarity path using cosine distance over sentence embeddings, and a sparse n-gram keyword path performing fuzzy substring matching over a relational keyword index. Results from both paths are merged with vector-matched entries taking precedence, producing a unified set of contextually relevant historical memory fragments. Memory entries are generated via an LLM-driven topic compression pipeline that distills entire conversation segments into structured JSON summaries (L1 topic summary, L2 one-sentence précis, and keyword sets), triggered automatically by the same LLM's topic boundary detection (new_topic flag). The system further serves as an adaptive fallback routing mechanism: when the primary LLM response indicates that user intent is not understandable, the system retrieves associative memory and invokes a secondary LLM call augmented with the retrieved context. We describe the architecture, retrieval algorithms, and integration within the iAware conversational AI framework, and discuss its implications for long-context dialogue management and multi-turn coherence.