MirrorMind: Empowering OmniScientist with the Expert Perspectives and Collective Knowledge of Human Scientists

TL;DR

MirrorMind tackles the gap between isolated AI scientists and social, historical scientific knowledge by introducing a hierarchical cognitive architecture with three memory levels and a multi-agent orchestration. The Individual Level stores high-fidelity author memory (Episodic, Semantic, Persona), the Domain Level builds a structured concept graph from OpenAlex data, and the Interdisciplinary Level coordinates reasoning across domains. Through four tasks—AuthorQA, complementary ideas, interdisciplinary collaboration, and cross-domain problem solving—the framework demonstrates improved factual grounding, personalized reasoning, and robust cross-domain synthesis. This approach offers a path toward cognitively aware, collaborative, and scalable AI scientists capable of deep domain translation and interdisciplinary insight.

Abstract

The emergence of AI Scientists has demonstrated remarkable potential in automating scientific research. However, current approaches largely conceptualize scientific discovery as a solitary optimization or search process, overlooking that knowledge production is inherently a social and historical endeavor. Human scientific insight stems from two distinct yet interconnected sources. First is the individual cognitive trajectory, where a researcher's unique insight is shaped by their evolving research history and stylistic preferences; another is the collective disciplinary memory, where knowledge is sedimented into vast, interconnected networks of citations and concepts. Existing LLMs still struggle to represent these structured, high-fidelity cognitive and social contexts. To bridge this gap, we introduce MirrorMind, a hierarchical cognitive architecture that integrates dual-memory representations within a three-level framework. The Individual Level constructs high-fidelity cognitive models of individual researchers by capturing their episodic, semantic, and persona memories; the Domain Level maps collective knowledge into structured disciplinary concept graphs; and the Interdisciplinary Level that acts as an orthogonal orchestration engine. Crucially, our architecture separates memory storage from agentic execution, enabling AI scientist agents to flexibly access individual memories for unique perspectives or collective structures to reason. We evaluate MirrorMind across four comprehensive tasks, including author-level cognitive simulation, complementary reasoning, cross-disciplinary collaboration promotion, and multi-agent scientific problem solving. The results show that by integrating individual cognitive depth with collective disciplinary breadth, MirrorMind moves beyond simple fact retrieval toward structural, personalized, and insight-generating scientific reasoning.

Figures (4)

• Figure 1: The MirrorMind Architecture. The framework maps the conceptual foundations of scientific memory (Left) into a three-level technical system (Right). The Individual Level models specific cognitive trajectories via tri-component memory; the Domain Level encodes collective disciplinary memory via navigable concept graphs; and the Interdisciplinary Level orchestrates these components through task decomposition and knowledge integration.
• Figure 2: The Individual Level Architecture. The diagram illustrates (left) the tri-component memory system (Episodic, Semantic, Persona) and (right) the four-stage retrieval workflow that orchestrates them to produce a final, stylistically authentic answer.
• Figure 3: Overview of the Domain Level. This level constructs a Semantic Index from OpenAlex data. A Domain Expert Agent is then equipped with a specialized toolset to navigate the conceptual graph and provide expert insights.
• Figure 4: The Interdisciplinary Level's multi-agent workflow. A Coordination Agent receives a query, dispatches tasks to the Specialist Layer (Domain and Author Agents), and routes their outputs to the Review & Integration Layer for validation and synthesis into a final answer.