Mandol

An In-Memory Layered Memory System
for Long-Term Conversational Agents

Unified representation, efficient storage, and accurate retrieval of complex memory information — providing next-generation cognitive architecture for conversational agents.

Get Started View on GitHub

92.21%LoCoMo Accuracy

88.40%LongMemEval Accuracy

5.4×Retrieval Speedup

4.8×Insertion Speedup

What is Mandol?

Mandol is an in-memory, layered memory system for long-term conversational agents, with efficient and precise retrieval capabilities. It achieves unified representation, efficient storage, and accurate retrieval of complex memory information, providing theoretical foundations and technical solutions for next-generation agent cognitive architectures.

The system fuses key-value, vector, and graph indexing paradigms into a unified in-memory data structure, exposing a minimalist add() → holistic_retrieve() operational model. Its core innovation transforms the traditional "passive recall–rerank" retrieval paradigm into a proactive "Query-Aware Routing → quantitative denoising → high-quality context generation" paradigm.

Core Innovations

Three breakthroughs redefining how conversational agents remember

🏗️

Layered Memory Model

Theoretical

A layered theoretical memory model dividing the system into base, high-level, and intelligent query layers, with a structured semantic graph unifying complex multi-relational memory representations.

Structured semantic graph for unified representation of complex memory
Implicit semantic edges generated on demand for precision–flexibility balance
Bidirectional traceability linking base and high-level memories

💾

Unified In-Memory Storage

Architecture

A unified storage architecture based on in-memory semantic data structures, where co-designed SemanticMap and SemanticGraph fuse key-value, vector, and graph capabilities at the physical level.

Native fusion of KV storage, vector indexing, and graph structures
Atomic hybrid retrieval operators eliminate cross-store I/O bottlenecks
Active-memory / durable-storage synergy balancing performance and capacity

🔍

Intelligent Routing & Retrieval

Retrieval

A Query-Aware Routing and quantitative retrieval method transforming retrieval from passive recall–rerank into a proactive understand–filter–summarize paradigm.

Query-Aware Routing dynamically selects memory sources by intent
Two-stage quantitative denoising with conflict resolution
Token-constrained high-quality context generation without LLM in loop

Benchmark Performance

SOTA-level accuracy with lower token consumption on long-term conversational memory benchmarks

LoCoMo Accuracy (%) Comparison

GPT-4o-mini	Avg. Tok	Single	Multi	Temp.	Open	Overall
Mem0	1.0k	66.71	58.16	55.45	40.62	61.00
MemU	4.0k	72.77	62.41	33.96	46.88	61.15
MemOS	2.5k	81.45	69.15	72.27	60.42	75.87
Zep	1.4k	88.11	71.99	74.45	66.67	81.06
EverMemOS	2.5k	91.68	82.74	79.34	70.14	86.13
Mandol	2.0k	93.82	85.11	89.10	65.63	89.48

GPT-4.1-mini	Avg. Tok	Single	Multi	Temp.	Open	Overall
Mem0	1.0k	68.97	61.70	58.26	50.00	64.20
MemU	4.0k	74.91	72.34	43.61	54.17	66.67
MemOS	2.5k	85.37	79.43	75.08	64.58	80.76
Zep	1.4k	90.84	81.91	77.26	75.00	85.22
EverMemOS	2.3k	95.32	89.01	90.13	77.43	91.97
Mandol	1.9k	95.36	92.20	87.85	79.17	92.21

Mandol achieves 92.21% overall on LoCoMo with only 1.9k tokens — outperforming EverMemOS (91.97% / 2.3k) while using 17% fewer tokens, and surpassing Mem0 (64.20%) by 28 points.

LongMemEval Accuracy (%) Comparison

GPT-4o-mini	Avg. Tok	SS-Pref	SS-Asst	Temporal	Multi-S	Know. Upd.	SS-User	Overall
MemU	0.5k	76.70	19.60	17.30	42.10	41.00	67.10	38.40
Mem0	1.1k	90.00	26.78	72.18	63.15	66.67	82.86	66.40
Zep	1.6k	53.30	75.00	54.10	47.40	74.40	92.90	63.80
MemOS	1.4k	96.67	67.86	77.44	70.67	74.26	95.71	77.80
Mandol	2.1k	96.67	98.21	78.95	74.44	88.46	97.14	85.00

GPT-4.1-mini	Avg. Tok	SS-Pref	SS-Asst	Temporal	Multi-S	Know. Upd.	SS-User	Overall
EverMemOS	2.8k	93.33	85.71	77.44	73.68	89.74	97.14	83.00
Mandol	2.3k	96.67	98.21	87.22	77.44	89.74	98.57	88.40

Mandol achieves 88.40% overall on LongMemEval with 2.3k tokens — outperforming EverMemOS (83.00% / 2.8k) by 5.4 points while using 18% fewer tokens.

Quick Start

Three-step operational model: add → build → retrieve

# Install Mandol (zero mandatory dependencies)
pip install mandol
 
# Optional backends
pip install mandol[faiss]      # FAISS vector index acceleration
pip install mandol[neo4j]      # Neo4j graph database
pip install mandol[all]        # Install all optional deps

1. Install→2. Add Memories→3. Build & Query→4. Persist

Citation

If this work is helpful to your research, please cite our paper

mandol.bib

@article{mandol2026,
title   = {Mandol: An In-Memory Layered Memory System
for Long-Term Conversational Agents},
author  = {Yuhan Zhang and Zhiyuan Guo and Ziheng Zeng
and Wei Wang and Wentao Wu and Lijie Xu},
journal = {arXiv preprint arXiv:260x.xxxxx},
year    = {2026}
}

The paper is forthcoming. The full author list and arXiv link will be updated upon publication.