Explain LabNote’s continuity and audit boundaries

docs/WHY_LABNOTE.md

@@ -1,8 +1,29 @@
-# LabNote alongside context files and model memory
+# LabNote alongside context files, agent memory and observability
 
-LabNote keeps selected project continuity in visible files. It complements a
-model's context window and repository instruction files; it does not replace
-them.
+LabNote is a human-controlled project ledger for work that moves between AI
+assistants, coding agents, chats and people. It keeps selected project
+continuity in visible Markdown artifacts and structured JSON registry records.
+
+Its job is not to enlarge a model's context window, automatically remember
+everything, or trace every model call. Its job is to keep the project thread:
+the sources, handoffs, responses, reviews, decisions and next actions that
+people decide should travel forward.
+
+## The record is selective
+
+Not every message belongs in a durable project record. LabNote keeps the
+stepping stones that make the next piece of work understandable:
+
+- a source or incoming packet;
+- a response, contribution or draft;
+- a review or correction;
+- a decision or signoff; and
+- the next action or handoff.
+
+That makes the retained trail smaller and easier to inspect than a full
+transcript or a growing context blob. An AI can help prepare the record, but
+the workspace does not silently harvest conversations: the human decides what
+belongs and what needs review.
 
 ## Model context is not project continuity
 
@@ -11,26 +32,60 @@ by itself, decide what should remain important after the session, show the next
 tool why a decision was made, or create a selective record that a human can
 inspect.
 
-LabNote does not enlarge a model’s native memory or silently capture your
-conversations. A human or an AI session deliberately writes the record. That is
-why the trail can be checked, corrected, reviewed, and carried to another
-tool.
+LabNote does not enlarge a model's native memory. A human or an AI session
+deliberately writes the record, so the trail can be checked, corrected,
+reviewed and carried to another tool.
 
 Context length is useful. Project continuity is a separate job.
 
+## Agent memory is a different trade-off
+
+Some AI-memory systems automatically extract, compress, index and retrieve
+information across interactions. That can be useful when an agent needs
+automatic recall.
+
+LabNote takes a different route. Basic ledger use needs no LabNote background
+service, database or model API key, and it does not make an automatic memory
+store. It keeps the selected project record in the repository, where the
+people running the project can see and govern it.
+
+These approaches can coexist. Use automatic memory when automatic retrieval is
+the need; use LabNote when the project needs a deliberate, visible handoff and
+decision trail.
+
 ## Context files set local rules
 
-Files such as `AGENTS.md` or `CLAUDE.md` are useful ways to tell an AI about a
-repository: where important files are, how to run tests, and what local rules
+Files such as `AGENTS.md` or `CLAUDE.md` are useful ways to tell an AI about
+a repository: where important files are, how to run tests and what local rules
 apply.
 
 LabNote complements them. Its job is to route ongoing project work: where a
 session begins, what it should read, where it may leave work, how that work is
-reviewed, and when the session should stop and ask.
+reviewed and when the session should stop and ask.
 
 A context file tells an AI what kind of repository it is in. LabNote gives it a
 route through the work happening there.
 
+## An audit trail is not full observability
+
+AI-observability tools can trace prompts, model calls, tool calls, timing and
+token use. They answer runtime questions such as “what did this system call?”
+
+LabNote records a different layer: the project artifacts people choose to
+retain, and the review, decision and handoff around them. It is a
+project-level, human-controlled audit trail—not a claim to capture every model
+call or every action automatically.
+
+## Rails make the routine legible
+
+The rails do not make a model deterministic, smarter or infallible. They make
+routine coordination work clearer: a known entry, a bounded reading route,
+clear write targets and defined points to stop and ask.
+
+That means an incorrect contribution can remain visible as part of the record:
+it can be reviewed, corrected, rejected or superseded rather than quietly
+becoming unexamined “memory.”
+
 ## Use the smallest useful amount
 
 If a one-shot answer is enough, use the best tool available and get on with it.

locales/zh-CN/docs/WHY_LABNOTE.md

@@ -1,15 +1,37 @@
-# LabNote 与上下文文件和模型记忆
+# LabNote 与上下文文件、智能体记忆和可观测性
 
-LabNote 将经过选择的项目连续性保存在可见文件中。它与模型的上下文窗口和仓库指令文件互补，而不是取代它们。
+LabNote 是一个由人掌控的项目工作台账，适用于在 AI 助手、编程智能体、聊天会话和人之间流转的工作。它将经过选择的项目连续性保存在可见的 Markdown 成品和结构化 JSON 登记记录中。
+
+它的职责不是扩大模型的上下文窗口、自动记住一切，或追踪每一次模型调用。它的职责是保留项目脉络：人们决定要向前传递的来源、交接、回复、审阅、决策和下一步行动。
+
+## 记录是经过选择的
+
+并非每一条消息都应进入持久的项目记录。LabNote 保留的是让下一段工作仍然可理解的踏脚石：
+
+- 来源或进入的任务包；
+- 回复、贡献或草稿；
+- 审阅或修正；
+- 决策或签署；以及
+- 下一步行动或交接。
+
+与完整对话记录或不断膨胀的上下文块相比，这让保留的轨迹更小、更容易检查。AI 可以协助准备记录，但工作台账不会悄悄收集对话：由人决定什么应被保留、什么需要审阅。
 
 ## 模型上下文不等于项目连续性
 
 更长的上下文窗口可以帮助模型在一次会话中阅读更多内容。但它本身不会决定什么应在会话结束后继续重要，不会向下一个工具说明为什么做出了某个决定，也不会自动留下人类可以检查的选择性记录。
 
-LabNote 不会扩展模型原生记忆，也不会悄悄捕获你的对话。记录由人类或 AI 会话有意写入。因此，这条轨迹可以被检查、修正、审阅，并带到另一个工具中。
+LabNote 不会扩展模型原生记忆。记录由人类或 AI 会话有意写入，因此这条轨迹可以被检查、修正、审阅，并带到另一个工具中。
 
 上下文长度很有用。项目连续性是另一项工作。
 
+## 智能体记忆是一种不同的取舍
+
+有些 AI 记忆系统会跨交互自动提取、压缩、索引和检索信息。当智能体需要自动回忆时，这会很有用。
+
+LabNote 选择了另一条路径。基本台账使用不需要 LabNote 后台服务、数据库或模型 API 密钥，它也不会构建自动记忆库。它把经过选择的项目记录保存在仓库中，让运行项目的人可以看见并治理它。
+
+这些方法可以共存。需要自动检索时使用自动记忆；项目需要有意、可见的交接和决策轨迹时使用 LabNote。
+
 ## 上下文文件设定本地规则
 
 `AGENTS.md` 或 `CLAUDE.md` 之类的文件很适合向 AI 说明一个仓库：重要文件在哪里、怎样运行测试，以及有哪些本地规则。
@@ -18,6 +40,18 @@ LabNote 与它们互补。它的职责是路由持续进行的项目工作：会
 
 上下文文件告诉 AI 它身处什么样的仓库。LabNote 给它一条穿过正在进行的工作的路径。
 
+## 审计轨迹不等于完整可观测性
+
+AI 可观测性工具可以追踪提示、模型调用、工具调用、时序和 token 使用。它们回答的是运行时问题，例如：“这个系统调用了什么？”
+
+LabNote 记录的是另一层：人们选择保留的项目成品，以及围绕它们发生的审阅、决策和交接。它提供的是项目层面、由人掌控的审计轨迹；并不声称会自动捕获每一次模型调用或每一个动作。
+
+## 轨道让例行协作清晰可见
+
+这些轨道不会让模型变得确定、更聪明或不会出错。它们让例行协调工作更清楚：明确的入口、有限的阅读路径、清晰的写入位置，以及定义好的停下并询问的节点。
+
+这意味着不正确的贡献也可以作为记录的一部分保留：它可以被审阅、修正、拒绝或取代，而不是悄悄变成未经检查的“记忆”。
+
 ## 使用最小但足够的量
 
 如果一次性答案已经足够，就使用手边最好的工具并继续推进。如果普通文件夹已经足够，就使用普通文件夹。