AuthorsYuhan Zhang, Zhiyuan Guo, Ziheng Zeng et al.
TL;DR
Mandol uses an agglomerative hierarchical memory graph plus quantitative retrieval to reach 92.21% accuracy on LoCoMo with GPT-4.1-mini, 0.24 points above EverMemOS.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
Long-term conversational agents rely on heterogeneous vector and graph databases, causing fragmented memory and high cross-database I/O latency for hybrid queries.
Similarity-based RAG retrieval introduces noise, misses correlated clues, and lacks token budget control, leading to incomplete, conflicting, or redundant memories that degrade LLM accuracy and efficiency.
HOW IT WORKS
Mandol centers on a hierarchical memory model, SemanticMap, SemanticGraph, and a quantitative query mechanism to unify representation and retrieval.
You can think of Mandol as a layered memory system where RAM-like in-memory graphs sit above disk-like DuckDB persistence, coordinated by a smart query router.
This design lets Mandol construct traceable, denoised, token-constrained contexts that a plain context window or vanilla RAG pipeline cannot provide.
DIAGRAM
This diagram shows how Mandol routes a query, denoises candidates, resolves conflicts, and builds a token-constrained context.
DIAGRAM
This diagram shows how Mandol is evaluated across datasets, backbones, and metrics for accuracy and latency.
PROCESS
Hierarchical memory model
Mandol first organizes raw dialogues into the basic memory layer and agglomerates episodic, semantic, and emotional memories into the high-level abstract memory layer.
Agglomerative semantic data structure
Mandol stores memory units and relationships in SemanticMap and SemanticGraph, fusing key value, vector, and graph structures for unified hybrid retrieval.
Query adaptive routing
Mandol analyzes query intent, selects relevant memory spaces, runs BM25, SPLADE, and dense search, and expands subgraphs to recover multi hop evidence.
Quantitative denoising and token constrained context generation
Mandol applies MAD based denoising, conflict resolution with arbitration scores, and MMR based selection to build a compact high quality context under a token budget.
KEY CONTRIBUTIONS
Hierarchical memory model with structured semantic graphs
Mandol uniformly represents basic and abstract memories as structured semantic graphs, enabling traceable episodic, semantic, and emotional memories with source_unit_uids links.
Agglomerative SemanticMap and SemanticGraph
Mandol designs SemanticMap and SemanticGraph to fuse key value, vector, and graph structures, eliminating cross database I O and supporting atomic hybrid retrieval operators.
Quantitative retrieval mechanism
Mandol introduces query adaptive routing, MAD based denoising, conflict resolution with arbitration scores, and MMR token constrained context generation, improving LoCoMo accuracy to 92.21%.
RESULTS
LoCoMo Overall
92.21%
+0.24 over EverMemOS
LongMemEval Overall
88.40%
+5.40 over EverMemOS
Search mean latency
82.2 ms
5.4× faster than MemOS mean 440.5 ms at 10 QPS
Add mean latency
39.7 ms
4.8× faster than MemOS mean 191.9 ms at 10 QPS
On LoCoMo and LongMemEval, which test long term conversational recall, reasoning, and temporal tracking, Mandol achieves the highest overall accuracy among agent memory systems. These results show that Mandol’s hierarchical memory and quantitative retrieval improve both answer quality and system efficiency under realistic serving loads.
BENCHMARK
On LoCoMo and LongMemEval, which test long term conversational recall, reasoning, and temporal tracking, Mandol achieves the highest overall accuracy among agent memory systems. These results show that Mandol’s hierarchical memory and quantitative retrieval improve both answer quality and system efficiency under realistic serving loads.
BENCHMARK
Overall QA accuracy (%) on LoCoMo using GPT-4.1-mini as the backbone.
BENCHMARK
Overall QA accuracy (%) on LongMemEval using GPT-4.1-mini as the backbone.
KEY INSIGHT
Mandol uses a lightweight retrieval backend (Qwen3 Embedding 0.6B and bge reranker v2 m3) yet still beats EverMemOS on LongMemEval by 5.40 points.
This is surprising because EverMemOS relies on larger 4B retrieval models, suggesting that Mandol’s memory architecture and quantitative retrieval matter more than model size.
WHY IT MATTERS
Mandol unlocks unified, low latency, token efficient memory for long term agents, combining structured graphs with quantitative routing and denoising.
Builders can now deploy conversational agents that track evolving user states across sessions with higher accuracy and much lower retrieval and insertion latency on consumer hardware.
Peilin Feng, Suorong Yang, Soujanya Poria
arXiv 2026 · 2026
Σ-Mem maintains historical competence evidence matrices M_p, a peer relationship matrix G, a memory readout r_{p,t}, and residual steering δ_{p,t} to guide multi-agent trust decisions. On the mixed counterfactual benchmark, Σ-Mem raises Qwen3-0.6B accuracy at CF@90 from 46.22% with the base model to 71.10% with Σ w/ G.
Nikhil Verma
· 2026
Focus Agent adds start_focus, complete_focus, a persistent Knowledge block, and an optimized Persistent Bash plus String-Replace Editor scaffold to actively compress context during long software-engineering tasks. On five hard SWE-bench Lite instances against a Baseline ReAct agent, Focus Agent achieves 22.7% token reduction (14.9M → 11.5M) while matching 3/5 = 60% task success.
Xiaohui Zhang, Zequn Sun et al.
· 2026
ActMem transforms dialogue history into atomic facts via Memory Fact Extraction, groups them with Fact Clustering, links them through a Memory KG Construction module, and uses Counterfactual-based Retrieval and Reasoning for action-aware answers. On ActMemEval, ActMem reaches 76.52% QA accuracy with DeepSeek-V3, beating LightMem’s 63.97% by 12.55 points and NaiveRAG’s 61.54%.
Answers use this explainer on Memory Papers.
Checking…