CAS Proposes Mandol: A Memory System for AI Agents

机器之心 · wechat · 2026-07-05

Institutions including the Institute of Software, Chinese Academy of Sciences, have proposed Mandol, a condensed memory-native hierarchical memory system. It addresses issues in current Agent memory systems that rely on heterogeneous combinations of vector, graph, and relational databases, such as representation fragmentation, high cross-database query overhead, noisy passive RAG-style retrieval, and a lack of Token budget control. The core idea is to consolidate fragmented memories and heterogeneous storage into a unified memory-native architecture.

Through the co-design of a hierarchical memory model, unified memory semantic data structures, and intelligent quantized retrieval, Mandol achieves the best overall accuracy among open-source systems on two long-conversation memory benchmarks, LoCoMo and LongMemEval (92.21% and 88.40% respectively). At a concurrency of 10QPS, it averages an 82.2ms retrieval latency (a 5.4x speedup) and a 39.7ms insertion latency (a 4.8x speedup). It maintains lower latency than existing systems even on consumer-grade laptops, demonstrating strong potential for edge deployment.

Original post →

More from coding & agent

coding & agent channel →