Enhance System-2 Global Selection with Flexible Graph Traversal and Planning

Develop more flexible graph traversal and planning mechanisms for the System-2 Global Selection process in Mnemis’s hierarchical graph to go beyond the current top-down selection and improve deliberate, structured retrieval.

Background

Mnemis’s System-2 Global Selection currently performs a top-down, layer-by-layer traversal over a hierarchical graph of categories to retrieve structurally relevant memory items. This design aims to complement similarity-based retrieval by providing a global, deliberate selection mechanism.

The authors identify as an open direction the enhancement of global selection with more flexible traversal and planning mechanisms, suggesting the need for richer search strategies to improve coverage and reasoning over the hierarchy.

References

While the results are strong, several important directions remain open. In future work, we plan to support more data modalities and enhance global selection with more flexible graph traversal and planning mechanisms.

— Mnemis: Dual-Route Retrieval on Hierarchical Graphs for Long-Term LLM Memory  (2602.15313 - Tang et al., 17 Feb 2026) in Conclusion

Second, the RAG memory depends on the coverage and quality of indexed layout priors: weak retrieval may introduce irrelevant relations or bias scenes toward common arrangements. This limitation is orthogonal to our grow-and-repair framework, and could be addressed by adopting stronger retrieval, reranking, or adaptive memory-update mechanisms from future advances in RAG.

— ScenePilot: Grow-and-Repair Policy for Text-Driven 3D Indoor Scene Generation  (2608.30307 - Zhang et al., 31 Aug 2026) in Section ‘Limitations and Future Work’, Conclusion; Appendix, Section ‘Data Leakage and Limitations’