MEMORYBEAR RESEARCH · 2026

让 AI 在漫长对话中，
依然记得真正重要的事。

What matters should survive
the longest conversations.

MemoryBear 在两个长期记忆基准上接受全量评测：标准化长期记忆能力与复杂多会话压力测试。

MemoryBear was evaluated end-to-end on two long-term memory benchmarks: standardized abilities and complex multi-session stress tests.

查看双榜结果Explore results了解评测方法 →Read methodology →
LongMemEval95.00%LLM-judge accuracy
LoCoMo91.54%LLM-judge accuracy
2,486全量问题Full-set questions
100%Single-session user
98.72%Knowledge update
706Mean context · LME
TWO BENCHMARKS · ONE MEMORY SYSTEM

双重验证长期记忆能力

Two benchmarks. One durable memory layer.

LongMemEval 检验标准能力，LoCoMo 检验更长、更复杂、更接近持续交互的记忆表现。

LongMemEval tests standardized abilities. LoCoMo tests longer, noisier and more complex conversational memory.

REFERENCE SNAPSHOT

公开结果对比

Public result comparison

MemoryBear 结果来自本地全量评测；竞品数字为公开资料快照，仅作水位参考。

MemoryBear uses our full local run; competitor figures are public reference snapshots only.

LoCoMo overall comparison

* 除 MIRIX 外，所有回答均使用短答案提示词生成。

* All responses are generated using a short-answer prompt except MIRIX.

LongMemEval category comparison

* 不同底座模型、回答模型、Judge、检索预算与数据版本会显著影响分数。本表不代表统一环境下的严格复现排名。

* Base models, answerers, judges, retrieval budgets and dataset versions materially affect scores. This is not a controlled same-environment ranking.

METHODOLOGY

先公开方法，再讨论领先

Disclose the method before claiming the lead.

我们把可复核性放在榜单之前。以下是本轮 MemoryBear 结果所对应的评测边界。

Reproducibility comes before the leaderboard. These are the evaluation boundaries behind the results.

Evaluation setupMemoryBear v0.3.11
Answer modelqwen3.7-plus
Judgeqwen3.7-plus
Embeddingtext-embedding-v4
RetrievalTop-K 10
Reranking未启用Disabled
WORKED EXAMPLES

看看它是怎么想的

See how it thinks

每个回答背后，MemoryBear 都会召回相关记忆、整理证据，再作答。下面是真实样例。

Behind every answer, MemoryBear recalls relevant memories, assembles the evidence, then responds. These are real traces.

知识更新Knowledge Update✓ 答对✓ Correct0.46s
What was my personal best time in the charity 5K run?

召回的记忆（按相关度）

Recalled memories (by relevance)

ExtractedEntityscore 1.0
[2023-05-25T20:21:00Z] The user is trying to plan out their fitness schedule for the upcoming weeks after May 25, 2023.；[2023-05-25T20:21:00Z] The speaker who believes their next soccer game with coworkers is scheduled for June 8, 2023.；[2023-05-25T20:21:00Z]…
ExtractedEntityscore 0.996
[2023-05-27T10:20:00Z] A concrete goal the user aims to achieve during a charity running event.
ExtractedEntityscore 0.983
[2023-05-27T10:20:00Z] A concrete goal for the user to complete a charity 5K run faster than 25 minutes and 50 seconds, scheduled after May 27, 2023.
Statementscore 0.98
<statement> <content>The user hopes to beat their personal best time in the charity 5K run.</content> </statement>
MemoryBear 的回答
MemoryBear's Answer
25:50
标准答案
Gold Answer
25 minutes and 50 seconds (or 25:50)
时间推理Temporal Reasoning✓ 答对✓ Correct0.73s
What was the first issue I had with my new car after its first service?

召回的记忆（按相关度）

Recalled memories (by relevance)

ExtractedEntityscore 1.0
[2023-04-10T17:15:00Z] The speaker who has always wanted to visit Yellowstone National Park.；[2023-04-10T17:15:00Z] The speaker who thinks a scenic route to Yellowstone National Park would be amazing.；[2023-04-10T17:15:00Z] The speaker who is open to suggesti…
Statementscore 0.782
<statement> <content>The user got the user's car serviced for the first time on March 15, 2023, and had a great experience.</content> </statement>
MemorySummaryscore 0.766
<summary> <content>The user plans to get their car detailed soon and is seeking local recommendations. Their vehicle received its first service on March 15th, which they described as a highly positive experience.</content> </summary>
Statementscore 0.749
<statement> <content>The GPS issue on the user's car was resolved quickly.</content> </statement>
MemoryBear 的回答
MemoryBear's Answer
GPS system malfunction
标准答案
Gold Answer
GPS system not functioning correctly
TRANSPARENCY

我们也公开它答错的地方

Where it gets it wrong

透明地展示典型失败样例，是可信 benchmark 的一部分。

Publishing typical failure cases is part of a trustworthy benchmark.

跨会话检索Multi-Session Reasoning✗ 答错✗ Wrong1.13s
How many items of clothing do I need to pick up or return from a store?

召回的记忆（按相关度）

Recalled memories (by relevance)

ExtractedEntityscore 1.0
[2023-02-15T03:35:00Z] The speaker who needs help organizing their closet.；[2023-02-15T03:35:00Z] The speaker who still needs to pick up their navy blue blazer from dry cleaning.；[2023-02-15T03:35:00Z] The speaker who wore the navy blue blazer to a meeting in…
MemorySummaryscore 0.848
<summary> <content>The user plans to store winter clothes using boxes at home. They recently exchanged a pair of boots purchased from Zara on February 5th and still need to pick up the replacement pair. They are seeking tips on how to effectively track items…
MemorySummaryscore 0.832
<summary> <content>The user plans to store winter clothes using boxes at home. They exchanged a pair of boots purchased from Zara on February 5th and still need to pick up the replacement pair. They requested advice on methods to track items pending pickup or…
MemoryBear 的回答
MemoryBear's Answer
Two items: a replacement pair of boots from Zara and a navy blue blazer from the dry cleaner.
标准答案
Gold Answer
3
时间推理Temporal Reasoning✗ 答错✗ Wrong0.51s
How many months ago did I book the Airbnb in San Francisco?

召回的记忆（按相关度）

Recalled memories (by relevance)

ExtractedEntityscore 1.0
[2023-05-27T06:10:00Z] The speaker who plans to look into the recommended transportation options for their trip to San Francisco.；[2023-05-27T06:10:00Z] The speaker who is staying at an Airbnb in Haight-Ashbury.；[2023-05-27T06:10:00Z] The speaker who plans to…
Statementscore 0.867
<statement> <content>The user had to book the Airbnb three months in advance for the wedding.</content> </statement>
Statementscore 0.865
<statement> <content>The user plans to arrange transportation from SFO airport to their Airbnb in San Francisco.</content> </statement>
MemoryBear 的回答
MemoryBear's Answer
The information is not present.
标准答案
Gold Answer
Five months ago
EXPLORE MEMORYBEAR

访问 MemoryBear

Explore MemoryBear

通过官方渠道了解产品、查看项目源码，并获取完整的评测数据与结果。

Learn more about the product, browse the source code, and access the complete evaluation data and results through our official channels.

01

Hugging Face

查看完整评测数据、逐题结果、检索记忆与配置文件。

Browse complete evaluation data, per-question results, retrieved memories and configuration files.

02

GitHub

查看 MemoryBear 开源项目与最新代码。

Browse the MemoryBear open-source project and latest code.
03

RedBear AI

访问红熊AI官网，了解产品与团队信息。

Visit the RedBear AI website to learn about the product and team.
READING GUIDE

如何阅读这些数字

How to read the numbers

  1. LongMemEval 与 LoCoMo 的规模、难度和上下文长度不同，两项总分不应相互比较。
  2. 竞品结果来自公开页面或历史快照，不代表统一环境下的严格复现排名。
  3. 自动 Judge 存在边界误判可能；正式技术报告应同时提供人工抽检与逐题证据。
  1. LongMemEval and LoCoMo differ in scale, difficulty and context length; their headline scores are not directly comparable.
  2. Competitor figures come from public pages or historical snapshots, not controlled reruns.
  3. Automated judges can misclassify edge cases; formal reports should include human audits and per-question evidence.