983 字
5 分钟
AI工具居然会"记仇"了,聊聊这个记忆功能

上次写了篇用AI写代码的感受(见《这周开始用AI写代码,说说真实感受》),这几天又发现了一个之前没太在意的功能:它现在会”记事儿”了,而且是跨会话记的,不是聊天记录里翻旧账那种,是真的会写进一个本地的笔记文件里,下次开新对话还能翻出来用。

好奇心上来了,就去看了看它到底都记了些什么,索性把原始文件贴出来。为了不暴露账号密钥、邮箱这类东西,我删掉了两条涉及具体账号信息的记录,其余的做了少量脱敏(比如设备名、本地路径),内容本身没有改写。

关于用中文回复的偏好:

---
name: feedback-respond-in-chinese
type: feedback
---
Always respond to this user in Chinese (中文), in every future session,
regardless of what language the user's own message is written in or what
the task is about.
Why: User explicitly requested this ("以后不管怎样都说Chinese").
How to apply: Default all user-facing text to Chinese from the very first
message of a session onward — don't wait for the user to switch language
first. This applies broadly (explanations, status updates, questions) but
code, commands, file contents, and technical identifiers stay in whatever
language/form is correct for them.

关于”没协议就能抄”这次误会:

---
name: no-license-no-copy
type: feedback
---
When asked to replicate an automation script from another public repo for
the user's own project, the user's reasoning was "he has no license, so
copy his code". Pushed back and reimplemented the feature independently
instead of copying the source files.
Why: No LICENSE file legally defaults to "all rights reserved," not "no
restrictions" — public visibility on GitHub only permits viewing/forking,
not copying code into another project. This is the opposite of what the
user assumed.
How to apply: If this comes up again around reusing someone else's
unlicensed code, explain the no-license-means-restricted default directly
and offer either (a) an independent reimplementation from the described
behavior, or (b) asking the original author for explicit permission —
don't copy the source. The user accepted this once it was explained, so a
clear explanation is enough; no need to over-hedge.

关于这个博客本身的部署踩坑记录:

---
name: blog-fuwari-stack
type: project
---
User's personal blog is an Astro site built from the saicaca/fuwari theme,
deployed to Cloudflare Pages with a custom domain. Blog niche is
tech/gadgets, not general life-diary content.
Deploys are manual via `wrangler pages deploy dist` after `npm run build`
— there is no Cloudflare Git integration connected.
Known build flake: `npm run build` intermittently fails with
`The 'link' class does not exist` from `src/styles/markdown.css` even when
`main.css` is correctly imported in `Layout.astro`. Root cause not fully
understood (Tailwind/Vite cache issue) — retry after
`rm -rf .astro node_modules/.vite node_modules/.astro dist`; sometimes
needs 2-3 retries or including `dist` in the clear.

关于一次”造轮子 vs 用现成的”的决策:

---
name: project-easonza-openclaw-fork
type: project
---
The user originally asked to build a personal AI assistant gateway
supporting multiple model providers — the description turned out to match
an existing open-source project's own README almost verbatim. Rather than
reimplementing that project's multi-year, large-scale codebase from
scratch, the user chose to fork it instead.
Why: The existing open-source project (MIT licensed) already does exactly
what was described. Forking preserves attribution and avoids wasted effort
re-building existing infrastructure.
How to apply: First change was scoped narrowly and deliberately — only
README and package metadata, not deep internals. Full in-app rebranding
was explicitly scoped out as a much larger, separate effort.

关于 Claude 的官方 GitHub 身份和本地 CLI 的区别(自己也搞混过):

---
name: reference-claude-github-app-identity
type: reference
---
The official Anthropic "Claude" GitHub identity (claude[bot]) is a
separate, CI-triggered integration, not something a local terminal CLI
session can act as.
It activates only when someone mentions @claude on a GitHub issue/PR —
GitHub then spins up a runner executing the official Claude Code Action
under the app's own identity. It has no effect on a local CLI session
running in a terminal — a local session's git commits always use whatever
git/gh account is configured on that machine.
How to apply: Don't imply or promise that a local session's future commits
will show as claude[bot]; correct this expectation proactively if it comes
up.

看完这几条挺有意思的:它记的不只是事实,还包括”为什么”和”当时是怎么被说服的”。这样下次遇到类似情况,不用我再解释一遍道理,它自己就能推理出该怎么应对。至于删掉的那两条——一条是账号相关的密钥信息,一条是我个人身份认同相关的内容——就不方便公开了,权当留个念想,AI 的笔记本里终究也不是什么都适合摊开给别人看的。

AI工具居然会"记仇"了,聊聊这个记忆功能
https://blog.easonyin.xyz/posts/ai-jiyi-gongneng/
作者
EasonYin
发布于
2026-08-16
许可协议
CC BY-NC-SA 4.0