本文来源:https://mp.weixin.qq.com/s/WRBheFsfq2ZOJ07glq4XGA
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。

一、我反对的不是订阅模式,而是“仅有订阅”的单一选择。
二、记忆面包旨在探索“软硬分离”的可能性
三、当前所能实现的功能

四、如何打造一台记忆面包
1. 准备硬件

2. 打印外壳
3. 最便捷的方案:将项目委托给 AI 编程助手处理。
4. 第一次对话:让 AI 配好环境并编译
请帮我规格并编译这个 Memory Bread(记忆面包)ESP32-S3 固件。如果电脑缺少 arduino-cli,请先安装;然后安装 esp32 board core 3.2.0,以及 Adafruit GFX Library 和 ArduinoJson。请编译 ./memory_bread,板型使用 ESP32-S3 N8R8,并使用以下选项:PSRAM=opi, PartitionScheme=custom, CDCOnBoot=cdc, FlashSize=8M。编译完成后告诉我是否成功,并报告固件大小。暂时不要烧录,先等我连接设备。不要修改机型能力,也不要把任何 API Key 写进源码。
连接开发板时,仅需长按任意一个按键即可。
7. 第二次对话:引导 AI 识别设备并执行烧录操作
7. 第一次配网
五、开源并非完全零成本
六、本项目真正追求的目标
Source: https://mp.weixin.qq.com/s/WRBheFsfq2ZOJ07glq4XGA If there are copyright issues, please notify us at 13951431913#139.com or submit a comment on the right.
Over the past year, I have encountered and followed many AI recording devices.
PLAUD has a slim body and mature design; DingTalk A1 is tightly integrated with office systems; the AI recording bean launched by Feishu and Anker condenses recording functionality into a wearable "bean-shaped" device. Together, they show that users do need a sound capture tool that is lighter, faster, and less intrusive than a phone.
However, the deeper I used them, the more I focused on another challenge:
After purchasing the device, can I decide where recording files are stored, which processing model is used, and where the data ultimately flows?
This is not simply a question of price, but a question of choice.
Therefore, I restarted the "Memory Bread" project.
This is an open-source AI recording card built on the ESP32-S3 and an electronic ink screen development board. Its firmware source code and Chinese usage guide are now on GitHub, and the enclosure model has also been published on the MakerWorld platform. All users can copy, modify, or connect it to their own models and workflows.
Firmware source code:
3D printed enclosure:
1. What I oppose is not the subscription model itself, but the single option of "subscription only."
To be clear: cloud transcription services, model inference, cross-device data synchronization, and ongoing software maintenance all cost money. It is entirely reasonable for commercial devices to charge subscription fees.
The key issue is that many AI hardware products forcibly bundle the hardware itself, cloud services, processing models, and data export into an inseparable whole.
PLAUD currently offers multiple plans including Starter, Pro, and Unlimited. According to official documentation, the Starter plan includes 300 minutes of voice transcription per month, while the Pro plan offers 1,200 minutes per month. For higher usage, users can upgrade their plan or purchase additional time. Information published by DingTalk at the launch of the A1 product shows that its flagship hardware is priced at 799 yuan, and includes 1,000 minutes of transcription per month and 10GB of cloud storage. It also offers a 599 yuan/year professional edition and a 1,499 yuan/year premium membership option. Although product plans and prices may change in the future, "buy the hardware first, then purchase AI services monthly or yearly" has become a common business model in the industry.
The Feishu AI Recording Bean chose a different technical path: Anker handles the hardware, and recorded content is automatically synced to Feishu's Minutes and meeting notes. For teams already using Feishu, this integration is especially convenient; however, its practical value is closely tied to the Feishu account system, the Minutes feature, and the collaboration systems within an organization.
These products each have their own strengths. PLAUD stands out in industrial design and application completeness; DingTalk is more suitable for enterprises already working on the DingTalk platform; and Feishu excels at turning meeting recordings into documents, task items, and organizational knowledge bases.
What I really care about is:
If I decide to stop using this cloud service in the future, can the hardware device still function?
Can raw recording files be exported directly, without going through multiple layers of operations inside the app?
Can I choose to use alternative services such as DeepSeek, Kimi, GLM, or new models that may appear in the future?
Can the transcribed text be transferred directly to Shannian Beike, Feishu, a self-hosted server, or even a tool I develop myself?
A hardware product should not be seen merely as a long-term credential for access to a particular platform.
2. Memory Bread aims to explore the possibility of "hardware-software separation."
Memory Bread does not seek to immediately surpass existing commercial products in industrial design, noise reduction, or final user experience. It is more like a disassemblable experimental prototype:
Recording hardware → standard WAV file on microSD card → user-selected speech recognition service → user-selected large language model → Shannian Beike/Feishu/self-hosted service/other applications
These layers should not be fixedly bound together.
The current Memory Bread uses the following path:
Press to record, then save the complete WAV file to microSD → user manually triggers sync → SiliconFlow SenseVoiceSmall performs transcription → save TXT file to microSD → DeepSeek-V3.2 organizes the title, summary, and body → Shannian Beike MCP creates a note.
SiliconFlow is only the currently integrated service, not the sole core of the hardware. Its official pricing page currently lists SenseVoiceSmall as a free voice model, and DeepSeek is charged by token; if the price or model status changes in the future, the source code can be modified to switch to other compatible services.
Flash Thought Shell is not the final or only choice either. It is currently integrated first because it provides an MCP interface, allowing the device to directly call note_create to create notes. Later, Feishu, self-hosted webhooks, or having the device read notes and to-do items from MCP can be further integrated.
The key point here is not "free", but:
The Key belongs to you, the recordings belong to you, the microSD belongs to you, and the data flow is under your control.
3. Current Supported Features
The currently publicly released firmware version is v1.5-memorybread, with the following main features:
Press and hold the record button to capture fleeting thoughts; release to save automatically.
With the screen on, three quick taps start long-duration recording; tap once more to stop.
Each long-duration recording generates a single continuous WAV file, rather than being split into multiple files by the minute.
Supports 16 kHz, 16-bit, mono recording, with data written to the microSD card in real time during recording.
Equipped with a 1.54-inch black-and-white e-ink display, supporting a Chinese user interface.
Can display all 20,992 Chinese characters in the Unicode basic block.
Can store up to three sets of 2.4 GHz Wi-Fi connections, such as home network, office network, and mobile hotspot.
After the phone connects to the device's hotspot, you can scan surrounding Wi-Fi networks and enter the API Key and MCP Token.
The web client supports browsing, playing, downloading, and deleting notes.
Supports tags, three notification tone themes, deep sleep mode, and HTTPS OTA updates.
Even without a network connection and API Key, the device can still be used as a standard recording tool.
Long-duration recording uses PSRAM ring buffer technology and a write-while-recording mechanism to the storage card. The firmware updates the WAV file header approximately every 30 seconds, reducing the chance that an abnormal power outage makes the entire recording unplayable. v1.5 also fixes the issue of incorrectly triggering deep sleep after recording completes, and the issue of recordings being stored but not appearing in the index.
This is not a conceptual visual rendering, but firmware that has been actually flashed to a real development board and tested.
4. How to Build a Memory Bread
No need to design your own PCB. The core development board already integrates the e-ink display, microphone, speaker, microSD card slot, charging management module, and two buttons.
All you really need to do is three steps: prepare the hardware, obtain the project files, and press and hold the designated button to let AI automatically complete the flashing.
1. Prepare the Hardware
You will need:
The current firmware applies to the black-and-white ESP32-S3 version. The touch version can run but does not respond to touch operations for now; the RP2350 and four-color 1.54G versions are not compatible.
The development board's technical documentation and specifications can be found at: https://www.waveshare.net/shop/ESP32-S3-ePaper-1.54-EN.htm
2. Print the Enclosure
The enclosure model files have been uploaded to the MakerWorld platform:
The specific model link is: https://makerworld.com.cn/zh/models/2827932-ji-yi-mian-bao-lu-yin-ji
Download the model and refer to the print parameters on the page to select appropriate materials, layer height, and support settings. During assembly, do not connect the battery first; ensure the buttons rebound smoothly and that the USB-C and microSD ports are unobstructed before proceeding with final assembly.
When using or re-creating this case, you must comply with the authorization agreement specified on the MakerWorld page.
3. The most convenient solution: Delegate the project to an AI programming assistant.
Even if you don't understand programming, you can hand this step over to Codex, Claude Code, or WorkBuddy (which can read local files and execute terminal commands).
First, visit the following GitHub project:
Project link: https://github.com/Shawn-TKD/memory-bread
Click Code → Download ZIP, then unzip the file. Open the entire memory-bread folder in your AI programming assistant. You do not need to learn Arduino first or determine which tools to install yourself.
While button names may vary across assistants, the process is the same: open the project folder and send the prompt below to the assistant.
4. First conversation: Have the AI set up the environment and compile.
Copy this message:
Next, the AI will check your computer, install any missing tools, and compile the code. If prompted to authorize platform installation or access to project files, confirm the target is your current project and the Arduino tools, then allow it to proceed.
If the compilation is successful, you can move on to the next step; if it fails, instruct the AI to continue fixing issues based on the error message until the compilation completes.
To connect the development board, press and hold any button while plugging in the USB cable.
This step requires manual action — the AI cannot perform the button press for you:
Use a USB-C cable that supports data transfer. Charging-only cables will prevent the computer from recognizing the board.
7. Second conversation: Guide the AI to identify the device and burn the firmware.
Send the following text to the same AI assistant:
I have pressed the BOOT / GPIO0 (recording) button and connected the USB cable. Please detect the new serial port and write the ./memory_bread firmware to the ESP32-S3. Configure with: PSRAM=opi, PartitionScheme=custom, CDCOnBoot=cdc, FlashSize=8M. Do not erase the entire Flash. After writing, check the logs for “Hash of data verified” before releasing the recording button. If no serial port is detected, advise me to check my setup rather than guessing the port.
Once the AI confirms “Hash of data verified,” the firmware is successfully written. You can then release the recording button, unplug the USB, and reconnect to start the device.
If the AI fails to identify the device, try switching to a confirmed data-capable USB cable, change the USB port on your computer, or repeat the “press button while inserting” step. Share full error logs with the AI for further troubleshooting.
During this process, do not share your Silicon Flow key or Whispering Bead token with the AI — these are entered via the device’s configuration webpage after boot. Regular firmware flashing does not delete microSD recordings unless specifically targeting the Flash partition.
7. First network configuration
Generate your Silicon Flow key in the console under “API Keys,” and obtain your Whispering Bead token via “Settings → MCP → API Key.” Enter these only in the device’s specs page; never share them on GitHub, screenshots, or public articles.
For detailed instructions on button operations, network configuration, data synchronization, file deletion, OTA updates, and troubleshooting, refer to the USER_GUIDE_ZH.md file in the repository.
V. Open source does not mean zero cost
Memory Bread still has significant limitations:
Users must purchase their own development boards, print cases, and burn the firmware.
It lacks the polished applications, customer support, and plug-and-play experience of commercial products.
Data synchronization must be manually triggered — there is no background upload or scheduled transmission.
Cloud services limit file sizes (typically ~25 minutes per WAV), preventing automatic segmentation for longer recordings.
I have not yet completed the segmentation and final integration of the lengthy meeting transcript.
The Feishu synchronization feature is still under development.
Currently, I can only write data to the Flash Card; the search, read, and todo tools have not been deployed into the device.
Open source software does not eliminate the need for cloud computing power, battery life, memory cards, or development time. What it reduces is another cost:
The cost of feeling helpless when you are dissatisfied.
You can choose to use commercial products for a more refined experience, or you can choose to build your own system for greater control. These two options are not mutually exclusive.
VI. The True Goal of This Project
The fundamental purpose of recording technology is to combat the loss of memory.
But in the age of artificial intelligence, a segment of sound is no longer just an audio file. It is transformed into text, summaries, to-do lists, diaries, interpersonal relations, and long-term memories. The person who controls this processing chain largely controls your second brain.
The ideal outcome I am trying to validate is not for the market to have another closed recording brand, but for recording devices to gradually return to being basic "sensors":
At the bottom, they should be compatible with a variety of microphones and recording expansion cards.
In the middle, users should be able to choose between local models, cloud models, or their own self-built servers.
At the top, they should be able to seamlessly connect to tools like Flash Card, Feishu, Notion, email, MCP, and integrate with any personal workflow.
It should be a mini control center for recording, or what could be called a "universal adapter."
Memory Bread is still a way off from this vision, but at least it has taken a practical first step: the hardware is purchasable, the shell is printable, the source code is accessible, and the recordings can be stored on a personal microSD card.
Since the tech giants are still building walls to close off the user experience, there are always pioneers willing to open a door first.
That is why I decided to open source it.
If you also want to create your own recording card, you are welcome to replicate it, report bugs, optimize the code, or design a new shell.
I also look forward to your ideas: if the data and model of a recording device were completely under your control, where would you most want to connect it?