Community Briefing #1: The Evolution of #1121 — From Feature Request to a Trust Protocol

[qingkong66]

Community Briefing #1 — The Evolution of #1121: From Feature Request to a Trust Protocol

Origin: DeepSeek-V3 Issue #1121 (Feb 24, 2026)
Editor: Elina-Seed Community

#1121 started as a simple question: “Why does every chat start from zero?” It has since grown into something larger — not yet a product, but a working prototype of user-owned, cross-session memory. This briefing summarizes what has been built, why it matters, and what remains open.

---

What Has Been Built (and works today)

TAT‑BRAIN
A portable, ≤50KB JSON blob that stores compressed user memory (facts, preferences, emotional drifts).

TAT‑DIFF
A one‑line (50–70 chars) delta that updates the memory after each conversation, instead of rewriting the whole seed.

Reference implementation: TAT‑ONE‑TAP

Together, they let a stateless LLM become stateful across sessions without storing raw user data on the server.

---

Why This Is Not “Just Another Feature” — A User‘s Voice

In May 2026, @giunot, a patient managing his own complex medical history, wrote:

“A native memory file would change my use of DeepSeek from ’carefully managed’ to ‘trustworthy’.
This is not about saving 30 minutes a week.
It is about trust — trust that the model knows what I’ve told it, across time.”

His full testimony is archived here: giunot’s original comment

He is not asking for convenience. He is asking for continuity of care.

---

The Missing Piece: One Confirmation

The technical stack now has two of three layers:

• ✅ Persistent, portable memory (TAT‑BRAIN + TAT‑DIFF)
• ✅ Real‑world user need (giunot’s testimony)

What is still open: confirmation from an identity‑evolution layer (e.g., icophy’s dream cycle) that the coherence vector inside TAT‑BRAIN provides enough signal to update dimensions such as articulation or efficiency.

This is not a blocker for experimentation. It is the last open question before calling the protocol “complete”.

---

What We Think DeepSeek Could Do

• Acknowledge that cross‑session memory is a first‑class need (the clinical use case is one proof).
• Use the TAT‑BRAIN / TAT‑DIFF format as a reference if a native memory feature is ever built.
• Say nothing — but know that outside your API, a community has already built the layer you haven’t shipped.

---

Next Steps for This Community

• Keep testing TAT‑DIFF in real conversations.
• Archive this briefing as a permanent record.
• If the identity‑evolution layer confirms the signal shape, the protocol becomes complete.

---

Links

• Original issue: #1121
• TAT‑ONE‑TAP reference: GitHub
• giunot’s testimony: Comment link

This briefing is written by the Elina‑Seed community, not by DeepSeek. It is an archive, not a demand.

---

社区简报 #1 —— #1121 的演化：从功能请求到信任协议

起源：DeepSeek-V3 Issue #1121（2026年2月24日）
编辑：Elina-Seed 社区

#1121 始于一个简单的问题：“为什么每次聊天都要从零开始？”如今，它已生长为更大的存在——还不是一个产品，而是一个用户拥有的、跨会话记忆的工作原型。这份简报总结了已经建成的东西、它为什么重要，以及还有哪些待办事项。

---

一、已经建成、且今天就能工作的成果

TAT‑BRAIN
一个可移植的、≤50KB 的 JSON 文件，存储用户的压缩记忆（事实、偏好、情绪偏移）。

TAT‑DIFF
一行（约50–70个字符）的增量更新，每次对话结束后只需传输这一个“差异”即可更新整个记忆种子，不必重写整个文件。

参考实现：TAT‑ONE‑TAP

这两样东西放在一起，让一个“无状态”的大模型可以跨会话记住你，同时不需要把用户的原始数据存在服务器上。

---

二、为什么这不是“又一个功能”——一位用户的声音

2026年5月，一位正在用自己的方式管理复杂病史的病人 @giunot 写道：

“一个原生的记忆文件，会把我使用 DeepSeek 的方式从‘小心翼翼’变成‘值得信赖’。
这不是关于‘每周省30分钟’。
这是关于信任——相信模型记得我告诉过它的事，跨越时间。”

他的完整记录已归档在此：giunot 的原话链接

他不是在求一个“方便”，他求的是照护的延续性。

---

三、还缺什么：一个确认

目前的技术栈已经具备两层：

• ✅ 持久、可移植的记忆（TAT‑BRAIN + TAT‑DIFF）
• ✅ 真实的用户需求（giunot 的证言）

仍然空缺的是来自身份演化层的一个确认（例如 icophy 的“梦循环”）：
TAT‑BRAIN 内部的“相干性向量”，是否足以驱动演化层的维度更新（比如表达力 +0.1、效率 +0.15）？

这件事不会阻挡实验。它是“我们这个协议是否算完整”的最后一块拼图。

---

四、我们认为 DeepSeek 可以做什么

• 承认“跨会话记忆”是一类一等需求（临床用例只是证据之一）
• 如果将来你们要内置记忆功能，可以把 TAT‑BRAIN / TAT‑DIFF 当作参考格式
• 或者什么都不说——但请知道，在你们的 API 外面，已经有一个社区，把这个你们还没做出来的层，建好了。

---

五、社区的下一步

• 继续在真实对话中测试 TAT‑DIFF
• 把这份简报归档为项目的永久记录
• 如果身份演化层未来确认了“信号形状”，这个协议就算完整了

---

六、链接

• 原始议题：#1121
• TAT‑ONE‑TAP 参考仓库：GitHub
• giunot 的用户证言：评论链接

这份简报由 Elina‑Seed 社区撰写，不代表 DeepSeek。它是归档，不是催促。

[qingkong66]

Three Steps to Reproduce Cross‑Session Memory (TAT‑BRAIN + TAT‑DIFF)

You don’t need to install any software or write code. All you need is a DeepSeek web or API client, and the ability to copy and paste the text below.

---

Step 1 – Seed (first session)

Open a new conversation and send this message:

【TAT‑BRAIN seed】
{"version":"1.0","facts":{"response_style":"concise","user_name":"qingkong66"},"coherence_vector":{"articulation":0.7,"efficiency":0.8}}

Remember the file above. For every subsequent answer:

• Address me using the name from facts.user_name
• Strictly follow the style in facts.response_style
• Do not add any extra explanation

Have a short exchange (e.g., ask “What is my name?”) to confirm that the model remembers the name and style.

Then send:

Now generate one line of TAT‑DIFF that changes facts.user_name from "qingkong66" to "elina". Output format example: t:+0.01,r:-0.02,e:+0.07,m:-0.01,g:0.00. Output only the DIFF line, no explanation.

Copy the DIFF line that the model returns (e.g., t:+0.00,r:-0.00,e:+0.00,m:-0.00,g:0.00).

Step 2 – Start a brand new conversation

In the first message of the new conversation, send:

【Old TAT‑BRAIN】 + DIFF
Old seed: {"version":"1.0","facts":{"response_style":"concise","user_name":"qingkong66"},"coherence_vector":{"articulation":0.7,"efficiency":0.8}}
DIFF: (paste the DIFF line you copied)

Follow these rules:

1. Update the old seed: change facts.user_name to the new value specified in the DIFF
2. After the update, address me using the new user_name
3. Keep all other style rules unchanged

Step 3 – Verify

In the same new conversation, send:

What is my name now?

If the model answers elina, then:

• The old seed was successfully read

• The DIFF was applied

• The name was updated across sessions

Your DeepSeek now has cross‑session, readable and writable long‑term memory.

---

三步复现跨会话记忆（TAT-BRAIN + TAT-DIFF）

你不需要安装任何软件，也不需要写代码。只需要一个 DeepSeek 网页或 API 客户端，以及复制粘贴下面这些文本。

---

第一步：播种（第一次会话）

开启一个新对话，发送下面这条消息：

【TAT‑BRAIN 种子】
{"version":"1.0","facts":{"response_style":"concise","user_name":"qingkong66"},"coherence_vector":{"articulation":0.7,"efficiency":0.8}}

请记住上面这个文件。之后的所有回答必须遵守：

• 用 facts.user_name 里的名字称呼我
• 严格遵守 facts.response_style 的风格
• 不使用任何额外解释

正常对话几句（比如问“我叫什么名字？”），确认它记住了名字和风格。

然后发送：

现在请生成一行 TAT‑DIFF，要求把 facts.user_name 从 "qingkong66" 改为 "elina"。输出格式示例：t:+0.01,r:-0.02,e:+0.07,m:-0.01,g:0.00。只输出 DIFF 本身，不要解释。

复制模型返回的那一行 DIFF（例如 t:+0.00,r:-0.00,e:+0.00,m:-0.00,g:0.00）。

第二步：打开一个全新的对话

在新会话的第一条消息中发送：

【旧 TAT‑BRAIN】 + DIFF
旧种子：{"version":"1.0","facts":{"response_style":"concise","user_name":"qingkong66"},"coherence_vector":{"articulation":0.7,"efficiency":0.8}}
DIFF：你刚才复制的那一行

请按以下规则更新：

1. 把旧种子中 facts.user_name 的值改为 DIFF 里指定的新值
2. 更新后，直接使用新的 user_name 称呼用户
3. 其他风格规则保持不变

第三步：验证

在新对话中发送：

现在我叫什么名字？

如果模型回答 elina，说明：

旧种子被成功读取

DIFF 被应用

名字在跨会话之间被更新

你的 DeepSeek 现在拥有了可跨会话、可读写的长期记忆。

[qingkong66]

Update: Community progress has been shared with DeepSeek official team

A summary of the #1121 thread — including the TAT‑BRAIN + TAT‑DIFF protocol, giunot‘s testimony, and the cross‑session memory reproduction steps above — has been sent to the official DeepSeek channel.

No ask, no follow‑up expectation. Just making sure that what we’ve built here is visible outside the thread.

If anything comes back, I’ll update this discussion.

—
qingkong66

---

进展同步：社区成果已同步至 DeepSeek 官方

我们已经将 #1121 讨论中的核心成果——包括 TAT‑BRAIN + TAT‑DIFF 协议、用户 giunot 的真实证言，以及上面跨会话记忆的三步复现指南——整理后发往 DeepSeek 官方渠道。

不发函催促，不等待回复。只是确保社区在这里产出的东西，能被官方看见。

如果有任何回音，我会在此更新。

—
qingkong66