Codex Memory Forge

Build a Codex that gets sharper over time.

Codex Memory Forge is a Codex-native memory system for people who are tired of re-explaining the same preferences, fixing the same mistakes, and rebuilding the same context every session.

No vector database. No daemon. No fake hidden memory. Just Codex, files, rules, and a clean refinement loop.

Instead of hiding memory behind a proprietary service, this project keeps the whole loop visible:

• AGENTS.md is the entrypoint
• layered markdown files hold memory
• promotion rules decide what becomes high-priority guidance
• an optional nightly automation keeps the system clean

Chinese guide

Why this project stands out

Most "self-improving agent" setups fall into one of two traps:

• they rely on platform-specific primitives that do not actually exist in Codex
• they turn memory into a messy dump that grows forever and gets worse over time

Codex Memory Forge takes the opposite approach:

• Codex-native, no fake hidden internals
• layered memory instead of one giant note
• explicit promotion rules instead of vague "the agent will remember"
• human-readable files you can audit, edit, back up, and version

What you get

• a reusable Codex skill
• a five-layer memory architecture
• a clean AGENTS.md entrypoint model
• a recommended nightly review automation
• templates for PROFILE.md, ACTIVE.md, LEARNINGS.md, ERRORS.md, and FEATURE_REQUESTS.md

Repo layout

.
|-- SKILL.md
|-- agents/
|   `-- openai.yaml
`-- references/
    |-- agents-snippet.md
    |-- memory-files.md
    `-- nightly-review.md

Quick start

1. Copy this project into your Codex skills directory.
2. Load the skill as $codex-memory-forge.
3. Create or update your global AGENTS.md using the guidance in references/agents-snippet.md.
4. Create the five memory files using references/memory-files.md.
5. Optionally add the nightly review automation using references/nightly-review.md.

The memory model

Use each file for a different class of signal:

• PROFILE.md: durable identity and communication preferences
• ACTIVE.md: short, high-priority operational rules
• LEARNINGS.md: reusable lessons not promoted yet
• ERRORS.md: recurring failures and debugging knowledge
• FEATURE_REQUESTS.md: long-term missing capabilities

This is the core idea of the project: memory should be curated, promoted, and pruned, not merely accumulated.

Who this is for

• engineers who use Codex every day
• solo builders who want Codex to adapt to personal workflow
• teams experimenting with durable agent behavior without extra infrastructure
• anyone who wants a transparent alternative to "trust me, the agent remembers"

Design principles

• visible over magical
• structured over bloated
• portable over locked-in
• operational over theoretical

Open-source direction

This repository is designed to be forked, audited, and improved in the open.

Recommended defaults if you publish your own version:

• keep the entrypoint thin
• keep the memory layers explicit
• keep promotion rules conservative
• never auto-edit AGENTS.md without clear user consent

License

MIT. See LICENSE.

If this project saves you repeated setup work, give it a star and ship your own Memory Forge.