Memory and learning

aivyx-coder carries knowledge between sessions in four ways, from the most deliberate to the most automatic.

`AGENTS.md` — instructions you write

Put conventions, architecture notes, build and test commands and “don’t touch X” rules in:

AGENTS.md at the project root — for this project;
~/.config/aivyx-coder/AGENTS.md — for every project.

Both are included in every turn and re-read each turn, so edits apply immediately. Your personal file comes first; where they conflict, the project’s wins. Each file has a token budget ([agents_file] budget_tokens, default 1024); a longer file is still included in full, with a one-time notice.

AGENTS.md is trusted: it goes straight into the model’s instructions. A project you’ve just cloned can therefore steer the model from the first turn — read its AGENTS.md like any other instructions before relying on it. It can’t get around the approval prompts or the sandbox.

Learned preferences

The model can propose additions to your personal AGENTS.md with its remember_preference tool — when you ask it to remember something, or when it notices a clear pattern. You see the diff and approve or refuse every time; this file never gets “always allow”. Turn it off with enabled = false under [persona]. Not available in --auto.

Notes it keeps

For smaller facts, the model has memory_write, memory_read and memory_forget: many small notes under topics, either for this project (project:…) or global (global:…). Notes are only recalled when the model asks for them, never added automatically. Saving and forgetting ask for approval like any change (and “always allow” covers one exact topic). Each topic keeps at most 1,000 notes. Not available in --auto.

The project wiki

/wiki has the model write and maintain docs/wiki/ — one page per part of the workspace plus an architecture overview — with every page write approved as usual. Later, bare /wiki regenerates only the pages whose source changed since they were written; /wiki <page> regenerates one. The repository map lists the pages so the model can read the right one when it needs it. Each page is its own turn, so a configured [verification] command runs once per page — a slow test suite makes a first full /wiki slow.

Skills

aivyx-coder ships a small library of skills — step-by-step guidance for things like systematic debugging, writing a plan, or scoping a task. The model loads one with load_skill when it’s useful. Add your own in <skill-name>/SKILL.md folders under [skills] project_dir or user_dir; turn the library off with enabled = false under [skills].