Repo: https://github.com/sugarpoo/codex-learning-loop
Codex Learning Loop is a safe local learning loop for Codex on macOS that maintains guidance files future Codex sessions can read.
Codex Learning Loop gives Codex a file-based memory workflow without pretending to train the model. It captures reusable learning candidates, reviews them on a daily/catch-up schedule, and updates global guidance files under ~/.codex. It is designed for people who use AGENTS.md, automate local workflows, and want a conservative way to reduce repeated mistakes across sessions.
Codex Learning Loop is a small macOS project for Codex users who want reusable lessons to survive beyond one session. It writes learning candidates to a local inbox, runs a LaunchAgent-backed maintenance pass with catch-up handling, and promotes clear lessons into files such as GLOBAL_MEMORY.md, ERROR_PATTERNS.md, and AGENTS.md. It does not train or fine-tune Codex, scan the whole computer, or replace human review. It simply maintains local guidance files under ~/.codex that future Codex sessions can read.
I made Codex Learning Loop: a local macOS workflow that helps Codex keep reusable lessons across sessions.
It captures learning candidates, runs daily/catch-up maintenance, and updates guidance files like AGENTS.md, GLOBAL_MEMORY.md, and ERROR_PATTERNS.md.
It does not train the model. It keeps local files future Codex sessions can read.
https://github.com/sugarpoo/codex-learning-loop
I published Codex Learning Loop, a small macOS project for people using Codex in repeated development workflows.
The idea is simple: when Codex or a user notices a reusable lesson, it can be written as a learning candidate under ~/.codex. A daily LaunchAgent maintenance pass reviews candidates, handles catch-up after missed runs, and updates local guidance files such as AGENTS.md, GLOBAL_MEMORY.md, and ERROR_PATTERNS.md.
This is not model training or fine-tuning. It is a file-based memory workflow that keeps future Codex sessions more consistent while staying scoped to local Codex files.
Repo: https://github.com/sugarpoo/codex-learning-loop
Title: I built a local learning loop for Codex on macOS
I use Codex across sessions and wanted a safer way to keep reusable lessons around without pretending the model is self-training.
Codex Learning Loop is a local, file-based workflow under ~/.codex. It lets you capture learning candidates, then a daily/catch-up LaunchAgent maintenance pass can classify them into guidance files such as AGENTS.md, GLOBAL_MEMORY.md, and ERROR_PATTERNS.md.
It does not read all chats, scan the whole computer, or train the model. It maintains plain-text guidance files that future Codex sessions can read.
Repo: https://github.com/sugarpoo/codex-learning-loop
Title: Show HN: Codex Learning Loop, a local file-based memory workflow for Codex
I built Codex Learning Loop for macOS users who want Codex to carry reusable lessons across sessions through local guidance files.
The project captures learning candidates into ~/.codex/learning-inbox, runs a daily LaunchAgent maintenance job with catch-up handling, and updates files such as GLOBAL_MEMORY.md, ERROR_PATTERNS.md, and AGENTS.md.
It does not train or fine-tune the model. It keeps a conservative, local, reviewable file workflow that future Codex sessions can read.
Repo: https://github.com/sugarpoo/codex-learning-loop
I published Codex Learning Loop: a local macOS workflow that keeps reusable Codex lessons in files like AGENTS.md, GLOBAL_MEMORY.md, and ERROR_PATTERNS.md. It uses a daily/catch-up LaunchAgent and stays scoped to ~/.codex. It does not train the model. https://github.com/sugarpoo/codex-learning-loop
A safe local learning loop for Codex on macOS.
No. It does not train, fine-tune, or modify the model. It maintains local guidance files future Codex sessions can read.
No. The workflow is file-based. It processes learning candidates and Codex-owned guidance files under ~/.codex.
No. The maintenance workflow is scoped to ~/.codex and is designed to avoid unrelated files.
Manual editing works. This project adds a repeatable inbox, classification rubric, reports, archive, and daily/catch-up maintenance so useful lessons are easier to review and keep.
Short, factual, reusable lessons. Good candidates describe a repeated mistake, a safer workflow, or a durable preference. Do not include secrets, private logs, personal notes, or sensitive information.
The current automation is macOS-focused because it uses LaunchAgents. The file layout and scripts may still be useful as references for other systems.