GAAMA: Graph-Augmented Memory for AI Agents
A recent study presents GAAMA (Graph Augmented Associative Memory for Agents), a framework aimed at equipping AI agents with enduring long-term memory throughout various user interactions. Existing techniques, such as flat retrieval-augmented generation (RAG), fail to maintain the structural connections between memories, while entity-centric knowledge graphs encounter mega-hub issues in conversational data, which diminishes the flow of relevance. GAAMA develops a concept-mediated knowledge graph via a three-phase process: preserving verbatim episodes, extracting atomic facts and topic-level concept nodes using LLMs, and synthesizing advanced reflections. This graph features four types of nodes (episode, fact, reflection, concept) linked by five structural edge types, with concept nodes enabling cross-cutting traversal to mitigate the mega-hub effect. The paper can be found on arXiv with the identifier 2603.27910.
Key facts
- GAAMA stands for Graph Augmented Associative Memory for Agents
- It addresses limitations of flat RAG and entity-centric knowledge graphs
- Uses a three-step pipeline: episode preservation, LLM extraction, reflection synthesis
- Graph has four node types: episode, fact, reflection, concept
- Five structural edge types connect the nodes
- Concept nodes avoid mega-hub effects in conversational data
- Paper published on arXiv with ID 2603.27910
- Announce type is 'replace'
Entities
Institutions
- arXiv