Datasets:
File size: 7,345 Bytes
84d77cd | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 | ---
license: apache-2.0
pretty_name: MemoryBear Evaluation Results
language:
- en
tags:
- evaluation
- memory
- longmemeval
- locomo
- llm
task_categories:
- question-answering
size_categories:
- n<1K
---
# MemoryBear Evaluation Results
This dataset repository contains the evaluation results for **MemoryBear**, a next-generation AI memory system developed by [RedBear AI](https://www.redbearai.com).
MemoryBear's core breakthrough lies in moving beyond the limitations of traditional "static knowledge storage". Inspired by the cognitive mechanisms of biological brains, MemoryBear builds an intelligent knowledge-processing framework that spans the full lifecycle of **perception → extraction → association → forgetting**.
Unlike traditional memory tools that treat knowledge as static data to be retrieved, MemoryBear emulates the hippocampus's memory encoding, the neocortex's knowledge consolidation, and synaptic pruning-based forgetting — enabling knowledge to dynamically evolve with life-like properties. This shifts the relationship between AI and users from passive lookup to proactive cognitive assistance.
- **Website**: <https://www.redbearai.com/>
- **Code**: <https://github.com/SuanmoSuanyangTechnology/MemoryBear>
## Benchmarks
We evaluate on two widely used long-term conversational memory benchmarks:
- **LongMemEval** ([xiaowu0162/LongMemEval](https://github.com/xiaowu0162/LongMemEval)) — 500 questions probing five core long-term memory abilities of chat assistants (information extraction, multi-session reasoning, temporal reasoning, knowledge updates, and abstention) over long user–assistant interaction histories.
- **LoCoMo** ([snap-research/locomo](https://github.com/snap-research/locomo)) — 1,986 questions over 10 very long multi-session dialogues, covering single-hop, multi-hop, temporal-reasoning, open-domain, and adversarial questions.
## Repository Structure
The evaluation artifacts are organized by benchmark (`lme/` for **LongMemEval**, `locomo/` for **LoCoMo**), then by system. Each run directory contains the same five artifacts:
| File | Description |
| --------------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| `*_metrics.json` | Aggregated metrics — accuracy, average context tokens, latency, and lexical scores (F1 / ROUGE / BLEU / METEOR), reported overall, by question category, and per question. |
| `*_hypotheses.json` | The answer generated by the system for each question, alongside the question, golden answer, and answer evidences. |
| `*_judged.json` | Per-question LLM-judge verdicts (`correct` / `incorrect`) together with the associated lexical metrics. |
| `*_retrieved_memories.json` | The memories retrieved by the system for each question, useful for inspecting retrieval quality. |
| `*_results.xlsx` | A spreadsheet summary of the run for convenient browsing. |
## Baseline Reproduction
The baseline results were reproduced by us using the official reproduction repos.
For comparability, the baselines were run under the same settings as MemoryBear: retrieval returns the **top 10** memories by default, and both hypothesis (answer) generation and LLM judging use **qwen3.7-plus** — identical to our own runs.
## Evaluation Results
### LongMemEval
Evaluated on the full **500-question** LongMemEval set. Accuracy is determined by an LLM judge.
| System | single-session-preference | single-session-assistant | temporal-reasoning | multi-session | knowledge-update | single-session-user | overall |
| ----------- | ------------------------- | ------------------------ | ------------------ | ------------- | ---------------- | ------------------- | --------- |
| MemoryBear | **100%** | 85.71% | **93.98%** | **93.98%** | **98.72%** | **100%** | **95.0%** |
| MemOS | 86.67% | **92.86%** | 81.95% | 80.45% | 94.87% | 98.57% | 87.4% |
| Memobase | 78.40% | 22.51% | 72.13% | 63.56% | 87.05% | 91.00% | 69.65% |
| Mem0 | 88.20% | 25.98% | 68.57% | 59.99% | 64.67% | 81.20% | 63.86% |
| Zep | 52.23% | 72.75% | 51.40% | 45.03% | 72.17% | 91.04% | 61.51% |
| Supermemory | 88.20% | 57.15% | 42.14% | 50.00% | 53.47% | 84.00% | 56.31% |
| MIRIX | 52.26% | 61.72% | 24.28% | 28.57% | 50.98% | 71.39% | 42.02% |
| MemU | 75.14% | 19.05% | 16.43% | 40.00% | 39.79% | 65.80% | 37.07% |
### LoCoMo
Evaluated on the full **LoCoMo** benchmark (1,986 questions across 10 conversations). Accuracy is determined by an LLM judge. The rest of the system baselines cover the 1,540 non-adversarial questions, so their `adversarial` cells are empty and their overall scores are computed over the remaining four categories.
| System | single-hop | multi-hop | temporal-reasoning | open-domain | adversarial | overall | overall F1 |
| ----------- | ---------- | ---------- | ------------------ | ----------- | ----------- | ---------- | ---------- |
| MemoryBear | **92.27%** | **90.78%** | **91.59%** | **73.96%** | 94.39% | **91.54%** | **67.49** |
| MemOS | 89.89% | 77.30% | 81.93% | 63.54% | – | 84.29% | 38.44 |
| Mem0 | 80.98% | 84.40% | 88.16% | 73.96% | – | 82.66% | 48.74 |
| Memobase | 71.66% | 61.42% | 77.14% | 51.53% | – | 69.68% | 50.18 |
| MIRIX | 66.86% | 51.55% | 65.11% | 45.47% | – | 62.29% | 28.10 |
| Zep | 64.91% | 49.51% | 52.08% | 32.33% | – | 57.39% | 41.23 |
| MemU | 65.01% | 59.96% | 25.75% | 48.50% | – | 54.87% | 35.15 |
| Supermemory | 65.95% | 48.56% | 30.18% | 41.39% | – | 53.72% | 34.87 |
<sub><i>\* The original LoCoMo dataset contains mislabeled golden answers. We corrected these mislabels, and all results above are reported on the corrected dataset.</i></sub>
|