A new memory architecture employs hyperbolic embeddings to organize vector database entries by an "abstraction score," allowing agents to manage knowledge more effectively than flat RAG systems. This approach structures data so that abstract entries, like preferences, sit near the center while concrete events reside at the boundary, facilitating better clustering and retrieval.
- Entries are assigned an abstraction score influencing HDBSCAN density-based clustering, visualized as gravitational "warp" in Poincaré space.
- Abstract entries have higher entropy and belong to multiple clusters, appearing frequently in retrievals alongside concrete details.
- The system returns "Spectrum Scenes," which are clusters containing a balanced mix of abstract, medium, and concrete entries centered on a theme.
- An agent processes these scenes to delete, consolidate, or abstract entries, preserving the hierarchical structure over time.
This method helps agents integrate new information into existing memory by providing extensive context through Spectrum Scenes, moving beyond coarse-grained knowledge graphs.