This paper aims to design a general scientific agent capable of performing scientific tasks to assist researchers in laboratory environments. Existing approaches have limitations, failing to meet the complexity and safety requirements of scientific tasks. Therefore, we present a novel agent, DAVIS, which integrates structural and temporal memory to enable model-based planning. DAVIS implements an agent-based multi-round retrieval system, analogous to human internal monologue, enhancing its ability to reason about past experiences. On the ScienceWorld benchmark, it outperforms existing approaches on eight out of nine scientific subjects and demonstrates competitive performance on the HotpotQA and MusiqueQA datasets. Notably, DAVIS is the first RAG agent to incorporate a conversational retrieval approach into the RAG pipeline.
