AuthorsZihao Tang, Xin Yu, Ziyu Xiao et al.
TL;DR
Mnemis uses dual-route retrieval with System-1 similarity search plus System-2 global selection on hierarchical graphs to reach 93.9 on LoCoMo and 91.6 on LongMemEval-S.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
Existing RAG and Graph-RAG systems rely almost entirely on similarity-based retrieval, which struggles with global reasoning and comprehensive coverage of relevant information.
When queries like “Which cities did Dave travel to in 2023?” require scanning dispersed facts, similarity search misses weakly related items, causing long-term agents to answer incompletely or incorrectly.
HOW IT WORKS
Mnemis combines a Base Graph of Episodes, Entities, Edges, and Episodic Edges with a multi-layer Hierarchical Graph plus System-1 Similarity Search and System-2 Global Selection.
You can think of Mnemis like RAM plus an indexed card catalog: System-1 is fast fuzzy lookup, while System-2 walks a semantic index of categories from coarse to fine.
This dual-route design lets Mnemis deliberately traverse semantic hierarchies to find structurally relevant but semantically distant memories that a plain context window or vanilla RAG would never surface.
DIAGRAM
This diagram shows how Mnemis processes a user query through System-1 Similarity Search and System-2 Global Selection to build the final memory context.
DIAGRAM
This diagram shows how Mnemis is evaluated on LoCoMo and LongMemEval-S, including ablations over System-1 and System-2 routes and backend models.
PROCESS
Mnemis Base Graph Ingestion
Mnemis converts new conversation turns into Episodes, extracts Entities, Edges, and Episodic Edges, and de duplicates them using embeddings and full text search.
Mnemis Hierarchical Graph Ingestion
Mnemis builds multi layer Categories over Entities using Minimum Concept Abstraction, Many to Many Mapping, and the Compression Efficiency Constraint.
System 1 Similarity Search
Given a query, Mnemis runs embedding search and BM25 over Episodes, Entities, and Edges, then merges rankings via Reciprocal Rank Fusion into top k results.
System 2 Global Selection
Mnemis traverses the Hierarchical Graph top down, selects relevant Categories and entities, retrieves connected Episodes and Edges, and re ranks the union for the answer.
KEY CONTRIBUTIONS
System 1 Similarity Search plus System 2 Global Selection
Mnemis introduces dual route retrieval that merges fast similarity search with deliberate top down Global Selection, reaching 93.9 overall on LoCoMo with GPT 4.1 mini.
Refined Base Graph Extraction
Mnemis improves Base Graph ingestion with richer entity fields, reflection based extraction, and forced speaker entities to enhance recall and robustness for long term memory.
Hierarchical Graph with Compression Efficiency Constraint
Mnemis constructs a Many to Many Hierarchical Graph guided by Minimum Concept Abstraction and layer wise Compression Efficiency Constraint to keep global selection scalable.
RESULTS
LoCoMo Overall
93.9 score
+1.6 over EverMemOS
LongMemEval S Overall
91.6 score
+9.6 over EverMemOS
LoCoMo Multi Hop
92.9 score
+1.8 over EverMemOS
LongMemEval S Multi Session
86.5 score
+8.0 over EverMemOS
On LoCoMo, which has about 600 turns per user and 1540 questions, Mnemis reaches 93.9 overall with GPT 4.1 mini, surpassing EverMemOS at 92.3. On LongMemEval S with roughly 115K tokens per session, Mnemis scores 91.6 overall, compared to EverMemOS at 82.0, showing that dual route retrieval scales to very long horizons.
BENCHMARK
On LoCoMo, which has about 600 turns per user and 1540 questions, Mnemis reaches 93.9 overall with GPT 4.1 mini, surpassing EverMemOS at 92.3. On LongMemEval S with roughly 115K tokens per session, Mnemis scores 91.6 overall, compared to EverMemOS at 82.0, showing that dual route retrieval scales to very long horizons.
BENCHMARK
LLM-as-a-Judge score on LoCoMo overall (Categories 1–4).
BENCHMARK
LLM-as-a-Judge score on LongMemEval-S overall.
KEY INSIGHT
System 2 Global Selection alone reaches 87.7 overall on LoCoMo, nearly matching or exceeding many Graph RAG baselines that rely only on similarity search.
This is surprising because structured top down traversal is slower and was expected to help only enumerative queries, yet Mnemis shows broad gains when combined with System 1.
WHY IT MATTERS
Mnemis unlocks deliberate, hierarchy aware memory retrieval where agents can scan semantic categories like Geography or Physical Health instead of guessing sub queries.
Builders can now design long lived agents that reliably answer multi hop and temporal questions over months of history without feeding 100K tokens into a single context window.
Guilin Zhang, Wei Jiang et al.
· 2026
A-MAC scores candidate memories using Utility, Confidence, Novelty, Recency, and Type Prior combined by a learned linear admission policy with Algorithm 1 A-MAC Memory Admission. On the LoCoMo benchmark, A-MAC achieves F1 0.583 and 2644 ms latency, improving F1 by 0.042 and reducing latency by 1187 ms compared to A-mem.
Robbyant Team, Zelin Gao et al.
arXiv 2026 · 2026
LingBot-World combines a Data Engine, Fundamental World Model, Action-Conditioned World Model, and Post-Training causal adaptation to turn a 28B-parameter video generator into a real-time interactive world simulator. On the VBench benchmark, LingBot-World achieves a dynamic degree of 0.8857 versus 0.7612 for Yume-1.5, while also improving imaging quality to 0.6683.
Manoj Madushanka Perera, Adnan Mahmood et al.
· 2026
AgenticAI-DialogGen chains ChatPreprocessor, KnowledgeExtractor, TopicAnalyzer, KnowledgeGraphBuilder, PersonaGenerator, DuelingChat Agent, ConversationValidator, ConversationRefiner, QAGeneration, and PostProcessing to turn raw multi-session chats into topic-guided, persona-grounded conversations with explicit short- and long-term memories. On the TGC / KG memory QA benchmark, Mistral-7B fine-tuned within AgenticAI-DialogGen achieves 87.36 F1, compared to GPT-4’s 83.77 F1 in a zero-shot setting on the same task.
Answers use this explainer on Memory Papers.
Checking…