Fast-Write, Deep-Read: EcphoryRAG as Dynamic Associative Memory for Lifelong Agents

Effective long-term memory is the cornerstone of autonomous agents capable of complex reasoning over extended horizons. However, current retrieval-augmented generation (RAG) systems face a critical trade-off: they are either computationally efficient but logically shallow (dense retrieval), or structurally rich but prohibitively expensive to update (knowledge graphs). Inspired by the cognitive neuroscience of memory \textit{ecphory}—where specific cues trigger the reconstruction of associative traces—we introduce \textbf{EcphoryRAG}, a neuro-symbolic memory architecture for agents. Unlike static graph methods that require heavy pre-computation, EcphoryRAG employs a "fast-write" mechanism, abstracting raw experiences into lightweight \textit{engrams} (entity-centric traces) with minimal latency. During retrieval, it utilizes a dynamic, centroid-based spreading activation algorithm to traverse implicit associations in vector space, simulating the brain's ability to recall multi-hop narratives from sparse signals. Extensive evaluations on the 2WikiMultiHop, HotpotQA, and MuSiQue benchmarks demonstrate that EcphoryRAG establishes a new state-of-the-art, improving the average Exact Match (EM) score to 0.475 while reducing memory consolidation costs by \textbf{18x} compared to leading graph-based baselines like HippoRAG2. These results validate EcphoryRAG as a scalable, high-fidelity cognitive substrate for the next generation of lifelong learning agents.