Multi-agent Systems
[Open soon]
Work description
The fellowship aims to investigate, develop and validate a knowledge graph-based approach to support the memory, organisation and visualisation of long conversations in multi-agent systems based on language models. The work will focus on the design of a conversational memory layer structured as a graph, capable of representing topics, entities, decisions, objectives, constraints, open questions and relationships between different parts of the conversation. This layer will be explored both as a context retrieval mechanism for agents with limited context windows and as a basis for new forms of interaction and visualisation of the evolution of the conversation, enabling users to understand, navigate and resume long and complex conversations more effectively. Memory architectures for conversational agents, techniques for the incremental construction of graphs from dialogue, mechanisms for retrieving relevant subgraphs to support response generation, and forms of interactive visualisation of the thematic and semantic structure of the conversation will be studied. The work will include an experimental and prototyping component, with evaluation in realistic long-conversation scenarios, measuring the quality of context retrieval, response consistency, reduction of dependency on the context window, usefulness of the visualisation, and user experience. Specifically, the main activities to be carried out by the fellowship holder are: • Study the state of the art in multi-agent systems based on language models, memory mechanisms for conversational agents, GraphRAG, conversational knowledge graphs, and visual interfaces for exploring long conversations. - Define a conceptual model for representing conversations as a graph, including node types, relationships, temporal attributes, information provenance, and update mechanisms throughout the interaction. - Design a multi-agent architecture for the construction, maintenance and use of the conversational graph, including specialised agents for information extraction, graph updating, detection of relationships between topics, context retrieval, and response generation. - Develop graph-based short-term memory mechanisms capable of selecting relevant subgraphs and converting them into structured context for models with reduced context windows. - Explore graph-based forms of visualisation and interaction, allowing users to navigate topics, decisions, open questions, dependencies and semantic relationships between different moments of the conversation. - Create a functional prototype that integrates the conversational graph layer with a multi-agent system based on language models. - Evaluate the prototype in long and complex conversation scenarios, comparing the proposed approach with alternative strategies such as sliding-window context, recurrent summarisation, or simple vector retrieval. - Analyse the results obtained in terms of response quality, conversational consistency, ability to retrieve older information, efficiency in the use of context, clarity of the visualisation, and perceived usefulness for the user. - Systematise the results of the work, producing technical documentation, evaluation reports and, where applicable, scientific contributions in the form of papers, demonstrations or reusable resources.
Academic Qualifications
Bachelor’s degree in Computer Engineering, Information Systems, or a related field.
Minimum profile required
Bachelor’s degree final grade above 12.
Preference factors
- Fluency in Portuguese and english. - Experience in the development and evaluation of memory management approaches in multi-agent systems.
Application Period
Since 14 May 2026 to 27 May 2026
[Open soon]
Centre
Industrial & Systems Engineering and Management