The Vision

The Vision | 愿景

AgentRecall is not a memory tool. It's the beginning of human understanding.

AgentRecall 不是记忆工具。它是理解人类的起点。

---

The Problem No One Talks About | 没人谈论的问题

Every AI memory tool stores PROJECT data — what was built, what was decided, what's blocked. That's necessary but insufficient.

每个 AI 记忆工具都存储项目数据——构建了什么、决定了什么、阻塞了什么。这是必要的，但不够。

The real problem: AI agents don't understand the HUMAN they're working with.

真正的问题是：AI agent 不理解与它们合作的人。

• The same agent gives the same response to a senior architect and a first-time coder

• A person who corrects an agent 50 times gets the same treatment on correction #51

• When a human says "do it", different humans mean entirely different things — but the agent can't tell

• 同一个 agent 给资深架构师和第一次写代码的新手同样的回复

• 一个人纠正 agent 50 次后，第 51 次仍然得到相同的对待

• 当一个人说"做吧"，不同的人意味着完全不同的事情——但 agent 分辨不出来

This is the Intelligent Distance problem — a structural gap between how humans think and how AI agents work. The gap can't be eliminated. But it can be navigated better with every interaction.

这就是智能距离问题——人类思维方式和 AI agent 工作方式之间的结构性鸿沟。这个鸿沟无法消除。但可以在每次交互中更好地穿越它。

---

The Goal | 目标

The same agent should perform differently for different people, and better over time for the same person.

同一个 agent 应该对不同的人表现不同，对同一个人随时间表现更好。

After 10 sessions, the agent should know:

10 次会话后，agent 应该知道：

What it learns	How it knows	如何知道的
This person communicates in scattered, non-linear bursts	Pattern from 8 alignment checks	8 次对齐检查的模式
When they say "save" they mean "git push"	Correction from session 3, confirmed 4x	第 3 次会话的纠正，确认 4 次
They care about results, not process	Feedback: "stop summarizing what you just did"	反馈："别再总结你刚做了什么"
When they're vague, it means they trust you	6 sessions of high-confidence alignment checks	6 次高信心对齐检查
When they're specific, it means they got burned before	Correction pattern: specific instructions follow mistakes	纠正模式：具体指令跟在错误之后

This isn't science fiction. It's pattern recognition on structured data that's already being collected.

这不是科幻。这是对已经在收集的结构化数据的模式识别。

---

Where AgentRecall Is Today | AgentRecall 目前在哪里

What works | 已经能做的

Project memory — AgentRecall reliably stores and recalls project context across sessions. Cold start, journal, palace, insights — these work and provide measurable value for multi-session projects.

项目记忆 —— AgentRecall 可靠地跨会话存储和回忆项目上下文。冷启动、日志、记忆宫殿、洞察——这些已经工作，并为多会话项目提供可衡量的价值。

Correction capture — alignment_check records every human correction. The data is stored permanently.

纠正捕获 —— alignment_check 记录每次人类纠正。数据被永久存储。

Insight compounding — the 200-line awareness cap forces merge-or-replace, so memory quality improves over time.

洞察复利 —— 200 行感知上限强制合并或替换，所以记忆质量随时间提升。

What's next | 下一步

Communication model — not just "what was said" but "how this person communicates." Scattered bursts vs. detailed specs. Prefers options vs. prefers recommendations. Learned, not manually coded.

沟通模型 —— 不仅是"说了什么"，而是"这个人如何沟通"。零散爆发 vs. 详细规格。偏好选项 vs. 偏好推荐。被学习，而不是手动编码。

Correction pattern analysis — aggregating correction data over time. After 20 corrections, the system should know: "60% of this person's corrections are about scope (agent does too much), 30% about priority (agent does the wrong thing first)."

纠正模式分析 —— 随时间聚合纠正数据。20 次纠正后，系统应该知道："这个人 60% 的纠正是关于范围的（agent 做太多），30% 是关于优先级的（agent 先做了错误的事）。"

Behavioral adaptation — accumulated understanding should shape the agent's approach BEFORE it starts reasoning about the task. Not just recall, but pre-conditioning.

行为适应 —— 累积的理解应该在 agent 开始推理任务之前就塑造它的方法。不仅是回忆，而是预调节。

---

The Bar | 标准

The question isn't "does it store data" — it does.

问题不是"它能存储数据吗"——它能。

The question is: Does the agent behave measurably differently on session 50 than session 1 with the same person?

问题是：与同一个人的第 50 次会话，agent 的行为是否可以衡量地不同于第 1 次？

That's the bar. Everything in AgentRecall is measured against it.

这是标准。AgentRecall 中的一切都以此为衡量。

---

For Contributors | 给贡献者

If you want to help build this:

如果你想帮助构建这个：

1. Alignment data is the goldmine. Every alignment_check is a labeled data point: what the agent thought, what the human meant, what the gap was. Aggregating these into behavioral priors is the highest-leverage work.

   对齐数据是金矿。 每次 alignment_check 都是一个标注数据点：agent 认为的、人类意图的、鸿沟是什么。将这些聚合成行为先验是最高杠杆的工作。

2. Fewer tools, deeper integration. 22 tools is too many for most agents to use effectively. The path forward is fewer, smarter tools that do more automatically.

   更少的工具，更深的集成。 22 个工具对大多数 agent 来说太多了。前进的方向是更少、更智能的工具，更多地自动完成。

3. Measure behavioral change. Build evals that compare agent behavior on session 1 vs. session N with the same user profile. If there's no measurable difference, the memory system isn't working.

   衡量行为变化。 构建评估，比较同一用户配置下第 1 次和第 N 次会话的 agent 行为。如果没有可衡量的差异，记忆系统就没有在工作。

---

See Also | 参见

• Intelligent Distance — the protocol in detail / 协议详解
• Core Concepts — how the memory layers work / 记忆层如何工作
• README — installation and quick start / 安装和快速入门