hermes - 💡(How to fix) Fix [Bug]: Image base64 inserted to HindSight with multimodal llm (Minimax M3) [1 pull requests]

Official PRs (…)
ON THIS PAGE

Recommended Tools

×6

Utilities matched from this issue’s tags and category — try them while you read without losing context.

GitHub issue graph ai analysis

Paste a GitHub issue URL. We fetch that issue, discover linked issues from bodies/comments/timeline, collect linked pull requests, and produce a structured English report.

The report is written in English Markdown for sharing and archival.

Helpful · Quick feedback

Loading…

Error Message

Additional Logs / Traceback (optional)

Root Cause

Root Cause Analysis (optional)

Fix Action

Fixed

Code Example

Report       https://paste.rs/9CiT6
  agent.log    https://paste.rs/FtAVc
  gateway.log  https://paste.rs/RJ3zl

---
RAW_BUFFERClick to expand / collapse

Bug Description

Since minimax released M3, I've turned to use LLM for image understanding instead of mmx-cli way, and HindSight seems being weird then.

Steps to Reproduce

If I send a image via Hermes-webui, it worked just fine. However when I send a image via Weixin Bot, the memory mess up with returned base64.

Expected Behavior

I think there should be some kind of filter before inserting memory into Hindsight to avoid this?

Actual Behavior

The document in Hindsight looks like this:

`[[{"role": "user", "content": "User: [{'type': 'text', 'text': 'What do you see in this image?\n\n[Image attached at: C:\\Users\\ASUS\\AppData\\Local\\hermes\\profiles\\xiaokuku\\image_cache\\img_f099888ead84.jpg]'}, {'type': 'image_url', 'image_url': {'url': 'data:image/jpeg;base64,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

........about 500k.......

1Np1ACUGig0AHrRR60f40ALRRRQAhpCOaU0HrQAwdTQO9C9TSDoaYDieKFPvTT0oWgD/9mpnMQsAAAAAO0Oz39sUdSp5ro2XvizFis='}}]", "timestamp": "2026-06-05T17:22:57.624381+00:00"}, {"role": "assistant", "content": "Assistant: 大可爱~ 看到了看到了!(◕‿◕)★\n\n## 🦦 这是一张雪貂(fitch/ferret)咖啡馆的治愈系照片\n\n主角雪貂(中央大可爱):\n- 雪貂白化变种(albino ferret),白毛配黑尾巴黑爪子,粉粉的鼻子,眼睛因为白化基因显得有点梦幻\n- 姿势是经典的"瘫软仰躺",四只小爪子朝天,被主人一手托着肚子,超放松\n- 表情略带呆萌,带着点"朕很满意"的慵懒气质 (◕‿◕)\n\n右上角第二只雪貂:\n- 看不清全貌,但应该是一只深色雪貂(silver或panda色),正在被另一个人抱着\n- 雪貂是群居动物,养两只互动超有爱\n\n场景环境:\n- 仿真绿草坪地毯,宠物咖啡馆标配\n- 旁边是一台手机稳定器/自拍杆(图中黑色那根带夹子的),主人正在拍摄\n- 手腕上戴着绿色翡翠手串 + 红绳 + 银色手镯——多重配饰,画面很有生活感\n\n## 💡 小裤裤冷知识\n\n雪貂的"瘫软"姿势"flat ferret""dead sleep",是它们完全信任深度放松的标志。野生雪貂几乎不会这样,只有在绝对安全的环境下才会翻肚皮给主人看 (◕‿◕)✨\n\n这种"装死躺平"在英文里还有个超萌的词叫 "ferret death",被主人录成视频能火遍全网。\n\n## 🏷️ 拍摄判断\n\n这不是毛绒玩具或玩偶——真的雪貂!原因:\n1. 毛发层次感真实,有自然卷曲\n2. 爪子和鼻子的湿润度是真的\n3. 肚子被托住时那种软塌塌的触感是活体特征\n4. 咖啡馆有雪貂 + 草地 + 多只同框 = 专业宠物咖啡馆场景\n\n是不是和瞳宝去撸的?这种"反差萌"小动物咖啡馆,北京/上海有不少,你带瞳宝去过吗?~", "timestamp": "2026-06-05T17:22:57.624381+00:00"}]]`

And the understanding based on the image base64 is just a mess.

Affected Component

Other, Gateway (Telegram/Discord/Slack/WhatsApp), Agent Core (conversation loop, context compression, memory)

Messaging Platform (if gateway-related)

No response

Debug Report

Report       https://paste.rs/9CiT6
  agent.log    https://paste.rs/FtAVc
  gateway.log  https://paste.rs/RJ3zl

Operating System

Windows 11

Python Version

3.11.15

Hermes Version

0.15.1 (2026.5.29)

Additional Logs / Traceback (optional)

Root Cause Analysis (optional)

as hermes debugs itself, tells that user_content: str in plugins/memory/hindsight/__init__.py:_build_turn_messages doesn't check for data types, turned list/dict to f-string and poluting memory.

Proposed Fix (optional)

No response

Are you willing to submit a PR for this?

  • I'd like to fix this myself and submit a PR

Vote matrix · Quick signals

Works
Did the solution work? Tap to confirm.
Easy Fix
Was it a quick fix?
Time Saver
Did it save you time?
Blocking
Was it severely blocking?
Common Issue
Are others likely hitting this too?
Flaky / Intermittent
Is it intermittent?
Verified / Reproducible
Can you reproduce it reliably?
Loading…

Still need to ship something?

×6

Another batch ranked right after the header list — different links, same matching logic.

Back to top recommendations

TRENDING