---
title: Where we are today
description: Codex and Claude Code keep useful memories, but their documented stores and controls differ.
section: Background
order: 3
---

**Documentation reviewed: November 12, 2026.** This comparison covers the memory systems described in the two official pages below. It does not test custom plugins or third-party services.

## Codex memories

Codex clients use a local memory store, separate from ChatGPT web memory. The documented locations are `$CODEX_HOME/memories/` and `$CODEX_HOME/memories_v2/`; the default Codex home is `~/.codex`. Store selection depends on configuration and migration state.

The documentation treats these files as generated state and recommends product controls for memory management. It does not document a configurable remote memory source or a common cross-harness record format.

**Remote memory: no documented remote source for local Codex memories.** This does not mean that ChatGPT has no cloud memory or that a custom integration cannot retrieve external data.

[Source: OpenAI — Memories](https://learn.chatgpt.com/docs/customization/memories?surface=app)

## Claude Code memory

Claude Code separates project instructions from auto memory. Its documented auto-memory location is `~/.claude/projects/<project>/memory/`, with a `MEMORY.md` index and topic files. `autoMemoryDirectory` can change the local directory.

The documentation states that auto memory is local to the machine and is not shared across machines or cloud environments.

**Remote memory: no documented remote auto-memory source.** Instruction files can be shared in a repository; that is distinct from a remote source for generated memories.

[Source: Anthropic — How Claude remembers your project](https://code.claude.com/docs/en/memory)

## The gap

| Property            | Local Codex memories               | Claude Code auto memory          | Open Memories proposal                      |
| ------------------- | ---------------------------------- | -------------------------------- | ------------------------------------------- |
| Location            | Codex home, version-specific store | Project-specific local directory | Explicitly configured local or HTTPS source |
| Organization        | Generated summaries and evidence   | Index and topic files            | Manifest and Markdown records               |
| Remote source       | Not documented on cited page       | Not documented on cited page     | Defined read contract                       |
| Cross-tool contract | Not defined on cited page          | Not defined on cited page        | Shared metadata, identity, and provenance   |

The proposal is to make those differences an adapter concern. The same test-command memory could then be read by more than one harness without rewriting the original record.

## Shared instructions are a separate layer

`AGENTS.md` and `CLAUDE.md` tell tools how to work. Memories capture context learned from work. A shared instruction file does not by itself define record identity, remote discovery, or memory updates.

Open Memories is intended to complement instruction files. Adoption still requires an adapter in each tool. No current Codex or Claude Code integration is provided by this site.