AGENTS.md, skills, and memory

Home Tools Platforms Codex AGENTS.md, skills, and memory

Separate checked-in project instructions from reusable procedures and locally recalled context.

Coding & development · Documented · Documentation checked 2026-10-04

Codex CLI, IDE extension, and desktop coding work; cloud environments need separate setup.

ChatGPT entitlement or supported API authentication, host capabilities, and workspace policy affect access.

The worked example is illustrative and has not been tested in a signed-in product account.

What it does and how it connects

Separate checked-in project instructions from reusable procedures and locally recalled context.

AGENTS.md provides deliberate repository guidance. Skills provide reusable task procedures. Local memory is an optional recall layer separate from ChatGPT web memory. Critical build commands and safety constraints should be explicit in checked-in instructions, not depend on whatever a previous conversation remembered.

Walkthrough

Use the product named in the access panel. The steps below combine documented controls with the editorial example.

  1. Inspect existing repository guidance before adding anything.
  2. Use the documented /init path or create a concise AGENTS.md with actual build and test commands.
  3. Enable and inspect local memory only where useful; do not store secrets.
  4. Test the instructions with an explanation-only repository task before using a reusable skill.

Worked example

A sample website repo can instruct Codex to use document links and preserve owner records. These are review criteria, not a substitute for database permissions. A skill can hold the repeatable review procedure for later tasks.

# Project instructions
Explain the existing content flow before editing.
Preserve owner records and unrelated working-tree changes.
Use the repository’s documented build and test commands.
Report actual verification results and blocked checks.
Do not merge, deploy, or alter external accounts without the task’s authorization.

Inspect and verify the result

Ask which instruction files were applied and compare them with the repo. Verify that named commands exist. Test a second task to ensure the procedure is reusable without relying on a hidden earlier conversation.

Limits and troubleshooting

Memory can update later or skip short sessions. ChatGPT projects do not automatically mount local repositories. If a rule must apply, supply it deliberately; if a skill is unavailable on a surface, use its approved instructions as explicit context.

Data, permissions, and actions

Instructions and memories may contain repository details. Review generated memory before sharing local state. Skills can bundle scripts, so inspect their code and references rather than trusting a task-oriented name.

Shell tools, MCP, and permissions

Distinguish a requested command from the filesystem, network, and external access actually allowed.

Explore Shell tools, MCP, and permissions

A scoped repository bug fix

Reproduce a sample defect and hand off a checked diff.

Explore A scoped repository bug fix

A prepared cloud task and review handoff

Run a bounded repository task in a reproducible cloud environment.

Explore A prepared cloud task and review handoff

Plugins, connectors, and skills

Separate packaged capabilities, reusable instructions, and the connections that let AI reach other software.

Explore Plugins, connectors, and skills

Sources and verification

Documentation checked 2026-10-04. Examples are illustrative unless described otherwise.

Projects and chats · OpenAI

Build skills · OpenAI

Memories across product surfaces · OpenAI

Share this page