修正編碼智能體兩次後會怎樣？Friday v0.2.6 的記憶功能

I build Friday, a free desktop app that runs coding agents like Claude Code, Codex and Cursor side by side on your own computer. Release v0.2.6 was about one thing I kept running into: you correct an agent, and the next one starts knowing nothing about it.

I started building Friday because I kept hitting Claude's usage limits and switching to Codex by hand. Every switch meant explaining the task again, and the corrections went with it. One of my Claude chats ran 12 days and auto-compacted 24 times without me clearing it once. You find out quickly what survives that and what doesn't.

What v0.2.6 added

When a run finishes, the coding tool that did it reads the run once more, on your computer, for what the next agent in that repository should know: a decision and why, a pitfall and its fix, a convention. For Claude Code that reading uses its small, fast model.

When you correct an agent, the run that fixes it is kept as a correction. Every agent that starts in that repository afterwards is told it first, whichever tool it is.

If the same mistake comes back after agents were told, Friday offers a command that would catch it, one that only reads or tests. Nothing runs until you add it. Once you do, Friday runs it after every change in that repository and sends a failure back to the agent, at most twice in a row. After that it's left to you.

A lesson no agent has needed in 60 days, or one a run found out of date, is retired, and you can bring it back.

What it costs, and what you can see

The reading is a request to your own coding tool, so it uses some of that tool's plan, and the run's text goes to that tool's provider like any other turn. It runs on your computer, through your own tool.

Everything it learned is in Settings › Memory, per repository. You can edit, pin or forget a lesson, add one yourself, or turn learning off. The lessons stay in Friday's folder on your computer, and they're per repository: nothing is shared between repositories or between people.

What I don't know yet

It's new, and I don't have numbers on how often it helps. An agent can write a wrong lesson, which is why all of them are visible and editable. The check is a command you add, and Friday only filters what it will accept; it isn't a sandbox, so read a command before you add it.

It's free, MIT-licensed, no account. It sends a few anonymous usage counts you can turn off, never your code or prompts. Source: https://github.com/splashxmoon/friday

If you try it, I'd like to hear what it learned that it shouldn't have.