moorcheh/memanto-longmem-results

Memanto — LongMemEval Benchmark Results

This dataset contains the full per-question evaluation results of Memanto on the LongMemEval long-term interactive memory benchmark, covering all five stages of a progressive ablation study. It accompanies the paper:

Memanto: Typed Semantic Memory with Information-Theoretic Retrieval for Long-Horizon Agents
Seyed Moein Abtahi, Rasa Rahnema, Hetkumar Patel, Neel Patel, Majid Fekri, Tara Khani
Moorcheh AI / EdgeAI Innovations
[arXiv preprint] · memanto.ai · moorcheh.ai

---

About Memanto

Memanto is a universal memory layer for agentic AI that achieves state-of-the-art accuracy on LongMemEval (89.8%) using a vector-only architecture — no knowledge graphs, no LLM-mediated ingestion, and no multi-query retrieval pipelines. It is built on Moorcheh's Information-Theoretic Search (ITS) engine, which provides deterministic semantic retrieval with sub-90ms latency and zero ingestion delay.

Key architectural features:

• 13-category typed semantic memory schema (fact, preference, decision, commitment, goal, event, instruction, relationship, context, learning, observation, error, artifact)
• Automated conflict resolution for contradiction detection and versioning
• Single retrieval query per question — no parallel or recursive strategies
• Zero ingestion cost — no indexing delay, no LLM extraction at write time

---

About LongMemEval

LongMemEval is a large-scale benchmark comprising 500 manually curated questions distributed across six categories, designed to evaluate five core memory abilities: information extraction, multi-session reasoning, temporal reasoning, knowledge update, and abstention. The standard evaluation setting spans approximately 115K tokens across ~50 sessions.

Category	Description
Single-session User	Information stated by the user within a single session
Single-session Assistant	Information from assistant responses within a session
Single-session Preference	User preferences expressed within a single session
Knowledge Update	Questions where earlier information was later corrected or updated
Temporal Reasoning	Questions requiring temporal ordering and reasoning
Multi-session	Questions requiring synthesis across multiple sessions

---

Dataset Structure

Each split corresponds to one stage of Memanto's five-stage progressive ablation study. Every row is a single evaluated question with the following fields:

Column	Type	Description
sample_id	string	Unique question identifier
question_type	string	LongMemEval category (single-session-user, single-session-assistant, single-session-preference, knowledge-update, temporal-reasoning, multi-session)
question	string	The evaluation question
prediction	string	Memanto's generated answer
ground_truth	string	Reference answer from the benchmark
score	int	Binary correctness score (1 = correct, 0 = incorrect), judged by Claude Sonnet 4
reasoning	string	LLM judge's reasoning for the assigned score
context_size	int	Number of tokens in the retrieval context for this question

---

Ablation Splits

The five splits map directly to the five stages of the ablation study reported in the paper:

Split	Stage	Configuration	LongMemEval Accuracy
ablation_1	Stage 1 — Naive Baseline	k=10, threshold=0.15, Claude Sonnet 4	56.6%
ablation_2	Stage 2 — Recall Expansion	k=40, threshold=0.10, Claude Sonnet 4	77.0%
ablation_3	Stage 3 — Prompt Optimization	k=40, threshold=0.10, optimized prompts	79.2%
ablation_4	Stage 4 — Maximum Recall	k=100, threshold=0.05, dynamic budget	85.0%
ablation_5	Stage 5 — Inference Model Upgrade	k=100, threshold=0.05, Gemini 3	89.8%

Key finding: Stage 2 (recall expansion, k=10→40) delivers the single largest gain (+20.4 pp), confirming that retrieval recall — not architectural complexity — is the dominant performance driver. The 19.6 pp gap between Stage 1's LoCoMo and LongMemEval results reflects LongMemEval's longer, multi-topic queries, which distribute semantic signal across a broader embedding space and are more sensitive to threshold calibration.

---

Final Results by Category (Stage 5)

Category	Accuracy
Single-session User	95.7%
Single-session Assistant	100.0%
Single-session Preference	93.3%
Knowledge Update	93.6%
Temporal Reasoning	88.0%
Multi-session	81.2%
Overall	89.8%

---

Comparison with Other Systems

System	LongMemEval	Architecture	Query Strategy
Memanto (ours)	89.8%	Vector Only	Single Query
Hindsight	91.4%	Hybrid (Reflection + Vector)	Parallel Multi-Query
EmergenceMem	86.0%	Hybrid (Graph + Vector)	Parallel Multi-Query
Supermemory	85.2%	Hybrid (Graph + Vector)	Parallel Multi-Query
Memobase	75.8%	Hybrid (Graph + Vector)	Parallel Single Query
Zep	71.2%	Hybrid (Graph + Vector)	Parallel Single Query
Full Context	60.2%	Full Context	N/A
Mem0	66.9%	Vector Only	Parallel Single Query
LangMem	58.1%	Vector Only RAG	Single Query

Memanto achieves the highest accuracy among all vector-only systems, surpassing Mem0 by +22.9 percentage points. Hindsight attains higher accuracy but requires maximum architectural complexity (complexity score 4/4): graph DB + LLM at ingestion + multi-query retrieval + recursive querying.

---

How to Load in Python

from datasets import load_dataset

# Load a specific ablation stage
ds = load_dataset("moorcheh/memanto-longmem-results", split="ablation_5")
df = ds.to_pandas()
print(df.head())

# Compare accuracy across all ablation stages
for stage in range(1, 6):
    ds = load_dataset("moorcheh/memanto-longmem-results", split=f"ablation_{stage}")
    df = ds.to_pandas()
    acc = df["score"].mean() * 100
    print(f"ablation_{stage}: {acc:.1f}%")

# Analyze per-category accuracy for Stage 5
ds = load_dataset("moorcheh/memanto-longmem-results", split="ablation_5")
df = ds.to_pandas()
print(df.groupby("question_type")["score"].mean().sort_values(ascending=False))

---

Evaluation Setup

• Benchmark: LongMemEval full 500-question suite (standard S setting, ~115K tokens, ~50 sessions)
• LLM Judge: Claude Sonnet 4 (all stages)
• Inference Model: Claude Sonnet 4 (Stages 1–4), Gemini 3 (Stage 5)
• Evaluation prompts: Adapted from the Hindsight repository to mitigate answerer refusal and rigid judge rejection of semantically correct but lexically divergent answers
• Memanto version: 2.1.4
• Retrieval backend: Moorcheh ITS engine (single query, no multi-query or recursive strategies)

---

Additional Assets

An Excel workbook (memanto_longmem_ablation_results.xlsx) containing all five ablation results as separate sheets with formatted headers is included in this repository for convenient offline exploration.

---

Reproducibility

All evaluation results in this dataset are fully reproducible. The moorcheh-ai/memanto-evaluation repository contains everything needed to replicate the LongMemEval experiments:

📁 memanto-evaluation/longmem — LongMemEval evaluation directory containing:

• 📊 The benchmark dataset and question suite
• 📝 Per-ablation result files (all 5 stages)
• 🐍 Python scripts to run the full evaluation pipeline

---

Citation

If you use this dataset or build on Memanto, please cite:

@misc{abtahi2026memantotypedsemanticmemory,
  title         = {Memanto: Typed Semantic Memory with Information-Theoretic Retrieval for Long-Horizon Agents},
  author        = {Seyed Moein Abtahi and Rasa Rahnema and Hetkumar Patel and Neel Patel and Majid Fekri and Tara Khani},
  year          = {2026},
  eprint        = {2604.22085},
  archivePrefix = {arXiv},
  primaryClass  = {cs.AI},
  url           = {https://arxiv.org/abs/2604.22085}
}

Also cite the LongMemEval benchmark:

@misc{wu2025longmemevalbenchmarkingchatassistants,
      title={LongMemEval: Benchmarking Chat Assistants on Long-Term Interactive Memory}, 
      author={Di Wu and Hongwei Wang and Wenhao Yu and Yuwei Zhang and Kai-Wei Chang and Dong Yu},
      year={2025},
      eprint={2410.10813},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2410.10813}, 
}

And the Moorcheh ITS engine:

@misc{abtahi2025hnswinformationtheoreticbinarizationrethinking,
      title={From HNSW to Information-Theoretic Binarization: Rethinking the Architecture of Scalable Vector Search}, 
      author={Seyed Moein Abtahi and Majid Fekri and Tara Khani and Akramul Azim},
      year={2025},
      eprint={2601.11557},
      archivePrefix={arXiv},
      primaryClass={cs.DB},
      url={https://arxiv.org/abs/2601.11557}, 
}

---

For questions, visit memanto.ai or moorcheh.ai.