GAAMA: Graph Augmented Associative Memory for Agents
Summary: arXiv:2603.27910v1 Announce Type: new
Abstract: AI agents that interact with users across multiple sessions require persistent long-term memory to maintain coherent, personalized behavior. Current approaches either rely on flat retrieval-augmented generation (RAG), which loses structural relationships between memories, or use memory compression and vector retrieval that cannot capture the associative structure of multi-session conversations. There are few graph-based techniques proposed in the literature, however, they still suffer from hub-dominated retrieval and poor hierarchical reasoning over evolving memory.
We propose GAAMA, a graph-augmented associative memory system that constructs a concept-mediated hierarchical knowledge graph through a three-step pipeline:
- Step 1: Verbatim episode preservation from raw conversations.
- Step 2: LLM-based extraction of atomic facts and topic-level concept nodes.
- Step 3: Synthesis of higher-order reflections.
The resulting graph uses four node types: episode, fact, reflection, and concept. These nodes are connected by five structural edge types, with concept nodes providing cross-cutting traversal paths that complement semantic similarity.
Retrieval combines cosine-similarity-based k-nearest neighbor search with edge-type-aware Personalized PageRank (PPR) through an additive scoring function. This innovative approach enables GAAMA to retrieve relevant information while maintaining structural relationships within the memory framework.
Performance Evaluation
On the LoCoMo-10 benchmark, which consists of 1,540 questions across 10 multi-session conversations, GAAMA achieves a mean reward of 78.9%. This performance significantly outperforms several existing models:
- Tuned RAG baseline: 75.0%
- HippoRAG: 69.9%
- A-Mem: 47.2%
- Nemori: 52.1%
Ablation analysis shows that augmenting graph-traversal-based ranking (Personalized PageRank) with semantic search consistently improves performance over pure semantic search on graph nodes, achieving an overall improvement of +1.0 percentage point.
Conclusion
GAAMA represents a significant advancement in the field of AI memory systems, particularly for agents that engage in multi-session interactions. By leveraging a hierarchical knowledge graph and innovative retrieval mechanisms, GAAMA not only enhances memory retention but also improves the coherence and relevance of responses. This development opens new avenues for creating more personalized and effective AI systems capable of understanding and retaining complex conversational contexts.