Mnemis: Dual-Route Retrieval on Hierarchical Graphs for Long-Term LLM Memory
Zihao Tang, Xin Yu*, Ziyu Xiao, Zengxuan Wen, Zelin Li, Jiaxi Zhou, Hualei Wang, Haohua Wang, Haizhen Huang, Denvy Deng, Feng Sun, Qi Zhang Microsoft ⋆ Corresponding Author · 2026
Read Mnemis: Dual-Route Retrieval on Hierarchical Graphs for Long-Term LLM Memory by Zihao Tang, Xin Yu*, Ziyu Xiao, Zengxuan Wen, Zelin Li, Jiaxi Zhou, Hualei Wang, Haohua Wang, Haizhen Huang, Denvy Deng, Feng Sun, Qi Zhang Microsoft ⋆ Corresponding Author (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2602-15313.txt.
Description
Tang and colleagues propose Mnemis, a recall setup joining System-1 likeness search with a partner System-2 route called Global Selection. AI recall, how models arrange and fetch old texts, grows more and more useful to large language models. Yet current tools like RAG and Graph-RAG mainly fetch recall through likeness routes. While fast, such System-1-style fetching fails with cases needing wide thinking or full reach of all needed info.
Mnemis arranges recall into a base graph for likeness fetching and a layered graph enabling top-down, careful walk over meaning layers. By joining partner powers from both fetching paths, Mnemis fetches recall bits that are both meaning and shape relevant. Mnemis hits state-of-the-art scores across all tested tools on long-term recall tests, hitting about 94 on LoCoMo and about 92 on LongMemEval-S using GPT-4.1-mini.
Catalog record
| CREATOR | Zihao Tang, Xin Yu*, Ziyu Xiao, Zengxuan Wen, Zelin Li, Jiaxi Zhou, Hualei Wang, Haohua Wang, Haizhen Huang, Denvy Deng, Feng Sun, Qi Zhang Microsoft ⋆ Corresponding Author |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 47 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2602-15313