AigentFlow Desktop is in pre-release. This documentation reflects the current build and will change before general availability.
Keeping control
Learnings
Short rules the app remembers about your machine and your project, so you are not correcting the same thing every session.
A learning is one sentence. "This project uses pnpm, do not use other package managers." It is stored, and it rides along with the first instruction of a session so the agent starts already knowing it.
Where they come from
| Source | How it happens |
|---|---|
| Failure detection | A command that was not found, a module that is not installed, a path that does not exist |
| Discovery scan | Reading the project: which package manager, which framework, which test runner, which formatter |
| You | Typing /learn |
Each one is filed under environment, tooling, pattern, preference or scope.
Teaching one directly
/learn always run the tests before saying it is done /learn --global never install a package without asking /memory list what has been learned for this project /forget pnpm remove learnings mentioning a keyword
Without the flag a learning belongs to the project. With it, the rule applies everywhere. A project is identified by its git remote, so a project without one saves globally.
Anything that looks like a credential is refused before it is stored. Passwords, API keys, tokens, private keys, connection strings and access keys never become learnings.
How they are used
Learnings are injected once per session, on the first instruction. They are ordered by how much they matter, environment first and personal preference last, and within a category the ones that have proved useful most often go first.
The number sent falls as the context window fills, from fifteen in a fresh session down to three in a full one. Teaching the same rule twice does not duplicate it, it counts as another use of the rule you already have.
Two hundred global learnings and one hundred per project are kept. Past that, the oldest go first.
Managing them
Settings has a searchable list, split into global and project learnings, showing how many times each has been used. A project learning can be promoted to global, any learning can be deleted, and the whole system can be switched off.