文档目录 · Contents
ZenStory AI ZenStory Workbench 工作台文档
AI 记忆与上下文
AI 记忆是项目的简短摘要，用来保留背景、风格和进度。详细人物、提纲与正文仍放在文件里，本轮要用什么就明确指出。
查看和编辑
在项目对话顶部打开“AI 记忆”，修改后点击保存。取消不会提交；重要更正后可以重新打开核对。
| 字段 | 适合记录 | 上限 |
|---|---|---|
| 项目简介 | 题材、主要处境、视角 | 4000 字符 |
| 写作风格 | 表达偏好与要避免的写法 | 1000 字符 |
| 当前阶段 | 已保存到哪里，下一项任务 | 1000 字符 |
| 写作约定 | 角色事实、情节安排、禁忌和长期创作边界 | 4000 字符 |
不用写满。把最关键的限制放前面，长篇背景放进对应文件。AI 可以更新简介、写作风格和当前阶段，发生重要决定后检查实际保存结果。写作约定由你维护：你在对话里明确说“记住……”时，AI 会把内容追加到写作约定，结果卡会列出这次记下的条目，并可直接打开对应文本框查看和编辑；其他时候 AI 不写写作约定，它的建议留在对话回复里。AI 只能追加，已有内容的修改和删除由你在 AI 记忆里完成。
区分已确认的事实和进度
把作者知道的真相、人物知道的信息和已经写出的事件分开。例如，《关窗之前》第1场只写到物件描述不符，可以这样记录：
项目简介：现实悬疑短篇，采用林禾有限视角。
写作风格：用问答和可观察的动作推进。
当前阶段：第1场已保存，包尚未交付；第2场待写。
写作约定：林禾不知道报错原因。沈敏与失主的关系、最终归属待决定。
依据文件：《林禾》《沈敏》《三场景提纲》《第1场 认领》。
不要把“林禾怀疑”改写成“林禾识破骗局”，也不要把下一场计划记录成已经完成。
更正错误记忆
先对照正文和人物资料，找到确实过时的字段，再修改那一处：
对照第1场与人物文件，检查当前 AI 记忆。
只列不一致的原句、依据和替换建议，先不修改文件或记忆。
采用后自己保存；简介、写作风格和当前阶段也可以明确让 AI 更新。记忆更正不会自动修好正文；剧情确实改变时，另行决定哪些文件需要修订。
换会话前记录进度
记录已经完成的工作、待定问题和下一步。新会话不会清除项目记忆，但也不会把所有旧聊天和整本书完整带入上下文。继续前打开目标文件，引用本轮关键内容。
AI Memory and context
AI Memory is a short project summary for background, style, and progress. Keep detailed characters, outlines, and prose in files, and identify what each task needs.
Read and edit memory
Open AI Memory at the top of project chat, edit, and save. Cancel does not submit changes. Reopen to check important corrections.
| Field | Useful contents | Limit |
|---|---|---|
| Project Summary | Genre, central situation, viewpoint | 4,000 characters |
| Writing Style | Expression preferences and approaches to avoid | 1,000 characters |
| Current Phase | What is saved and the next task | 1,000 characters |
| Writing Conventions | Character facts, plot commitments, prohibitions, and lasting creative boundaries | 4,000 characters |
You do not need to fill the limits. Put important constraints first and detailed background in files. The AI can update the summary, writing style, and current phase; check stored results after important decisions. Writing Conventions are yours: when you explicitly say "remember ..." in chat, the AI appends the content to Writing Conventions. The result card lists what it saved and opens the same field for review and editing. Otherwise the AI does not write Writing Conventions and keeps its suggestions in its replies. The AI can only append; edit or delete existing content yourself in AI Memory.
Separate decisions and progress
Separate author knowledge, character knowledge, and events already written. After Scene 1 of Before Closing reaches the mismatched item:
Summary: Realist mystery in Lin He's limited viewpoint.
Style: Advance through questions and observable actions.
Current Phase: Scene 1 saved; the bag remains at the desk. Scene 2 is unwritten.
Writing Conventions: Lin does not know why the description is wrong. Shen's relationship
with the owner and the final recipient remain undecided.
Evidence: Lin He, Shen Min, Three-scene outline, Scene 1 — The claim.
Do not turn suspicion into an exposed fraud or record the next scene's plan as completed.
Correct a mistaken memory
Compare prose and character notes, find the outdated field, and change that part:
Check AI Memory against Scene 1 and the character files.
List conflicting sentences, evidence, and replacements only.
Do not change files or memory yet.
Save the accepted correction yourself; you can also ask the AI to update the summary, writing style, or current phase. Correcting memory does not automatically repair prose. If the plot changes, decide separately which files need revision.
Record progress before a new chat
Record completed work, open decisions, and the next task. A new chat retains project memory, but it does not include every old message or the whole book in full. Open the target and quote key evidence before continuing.
Related: AI assistant and writing workflow.