Blog
Codex memories: how to turn them on, where they live, what they share
Published: October 28, 2026
Yes, Codex has memory. Codex memories are notes Codex writes for itself from your earlier chats and keeps on your machine under ~/.codex/memories/. They are off by default: turn them on with memories = true under [features] in config.toml, or in the ChatGPT desktop app under Settings, Personalization. They are separate from ChatGPT memory, they stay on one machine, and OpenAI says rules that must always apply belong in AGENTS.md.
Every fact below is from OpenAI's memories docs or the openai/codex repository, read on 28 October 2026, at Codex CLI 0.165.0. Memory has changed in most recent releases, so check the linked page before you trust a default.
Does Codex have memory like Claude Code?
Close to it. Both write their own notes from your sessions and keep them on your machine, apart from the instruction files you write yourself. Three things differ. Claude Code's auto memory is on by default and Codex memories are off. Claude Code keeps one memory folder per repository, while Codex keeps one store in your Codex home. And Codex does not write during a chat: it builds memories later, in the background, from chats that have gone idle. Our note on Claude Code auto memory covers the other side.
How do I turn Codex memories on or off?
In the CLI, set the feature flag in ~/.codex/config.toml:
[features] memories = true
| Where | Control | What it changes |
|---|---|---|
| config.toml | memories = true under [features] | Memories for this Codex home, off by default |
| ChatGPT desktop app | Settings, Personalization, Enable Codex memories | Memories for the selected machine |
| ChatGPT desktop app | Allow memories from tool-assisted chats | Whether chats that used MCP tools or web search can feed memories |
| CLI or app, in a chat | /memories | Whether this chat can use memories or become one. It does not change the global setting |
The IDE extension uses the store of the Codex host it connects to. To switch memories off, set the flag back to false or untick the setting.
Codex memories config: every setting
From the Codex config reference. All of them sit under [memories] in config.toml:
| Setting | Default | What it does |
|---|---|---|
generate_memories | true | When false, new chats are no longer stored as input for memories |
use_memories | true | When false, Codex stops adding existing memories to new sessions |
disable_on_external_context | false | When true, chats that used MCP tools, web search or tool search are kept out |
min_rollout_idle_hours | 6 | How long a chat must sit idle before it is considered |
max_rollout_age_days | 30 | The oldest chat considered |
max_unused_days | 30 | A memory not used for this long drops out of consolidation |
min_rate_limit_remaining_percent | 25 | Generation waits while less of your rate limit than this is left |
max_rollouts_per_startup | 16 | Chats processed per startup pass |
Where are Codex memories stored?
Under your Codex home, ~/.codex by default (or CODEX_HOME), in memories/. The docs say the folder holds “summaries, durable entries, recent inputs, and supporting evidence from prior chats.” The memories README in the Codex repo names the files: MEMORY.md, memory_summary.md, raw_memories.md, a rollout_summaries/ folder with one file per chat, and skills/. The folder is also a small git repository, so Codex can diff each pass against the last.
From 0.165.0 (27 October 2026), a profile that turns memories on for the first time gets the new v2 layout, which the code keeps in ~/.codex/memories_v2/ (PR #56211). Profiles that already had memories stay on v1.
OpenAI calls these files generated state: open them to see what Codex kept, but do not rely on editing them by hand. Secrets are redacted, not guaranteed absent, so read the folder before you share your Codex home with anyone.
When does Codex write memories?
Not when a chat ends. The README says the pipeline runs in the background when a new session starts (sub-agent sessions excluded). With the default settings it picks up chats that have been idle for six hours and are at most 30 days old, and the docs say it skips short ones. Each chat is summarised, then one pass merges the results. A pass is skipped while less than 25 percent of your rate limit is left.
So a correction you give Codex this morning is not in memory this afternoon. If it has to hold from the next session on, write it into AGENTS.md.
How do I clear or reset Codex memory?
- Wipe it. In the ChatGPT desktop app, Settings, Personalization, Delete Codex memories resets the selected machine's store, and clears Computer History on that machine too. The docs list no reset command for the CLI.
- Keep what is there, stop adding.
generate_memories = false. - Stop using it without deleting it.
use_memories = false. - One chat only.
/memoriesin that chat.
Does Codex share memory with ChatGPT?
No. In OpenAI's words, “ChatGPT web uses ChatGPT memory, while local Codex clients use a separate local memory store and controls.” What you tell ChatGPT in the browser does not reach Codex memories, and the reverse.
The exception is Computer History, a macOS feature of the ChatGPT desktop app that turns activity in the apps and sites you allow into memories both ChatGPT and Codex can use. It writes them as Markdown files under ~/.codex/memories/extensions/skysight/.
What happened to Codex Chronicle?
OpenAI's docs say “Computer History replaces the earlier Chronicle research preview, but it is a rebuilt system rather than a rename.” It is off by default, needs Codex memories on, and is for Pro, Business and Enterprise, where an admin grants access before anyone can opt in. It records clicks, typing, shortcuts and app switches, never screenshots or audio. The raw event files are deleted after 48 hours, and the memories it writes are not encrypted. OpenAI warns that it “increases the risk of prompt injection from content in apps and websites.”
Can Codex read Claude Code memory?
Only if you import it. /import in a local Codex CLI session copies a Claude Code or Cursor setup across, and the import docs map “Project memories from Claude Code” to Codex memories and instruction files to AGENTS.md. In the CLI it is a one-off copy. The desktop app can keep imported work in sync under Settings, Import.
Nothing flows the other way. Claude Code, Cursor and GitHub Copilot do not read ~/.codex/memories/, and each keeps its own (see Claude Code memory vs Codex, Cursor and Copilot, GitHub Copilot Memory and what happened to Cursor memories).
Are Codex memories shared with my team?
No. Memories are per machine, and the docs describe no team or repository scope. OpenAI's own advice is the line to keep: “Keep required team guidance in AGENTS.md or checked-in documentation,” and treat memories as “a helpful recall layer, not as the only source for rules that must always apply.”
| AGENTS.md | Codex memories | |
|---|---|---|
| Who writes it | You | Codex, in the background |
| Default | Loaded every session, in a trusted project | Off |
| Reaches teammates | Yes, through git | No |
| Reviewed before use | In the pull request | No; you can read the files after |
AGENTS.md has its own limits, and our note on why Codex ignores AGENTS.md covers them.
How teams fill the gap
- Commit the rule. When a correction matters past one session, write it into
AGENTS.mdin the pull request that needed it. - Share skills, not memories. A skill in the repo reaches everyone who opens it (see Codex skills).
- Use a shared memory. An MCP memory server or a company brain that every agent reads, so the rule is not stuck in one tool (see how to share memory and rules across tools).
Harbor is that last kind: shared memory for AI coding agents. It learns from where your team decides (Slack, pull request reviews, docs and agent sessions), lets you set what waits for a person, and gives Claude Code, Cursor, Codex and GitHub Copilot the rules that apply to each task.
Questions
Does Codex have memory?
Yes. Codex memories are notes Codex writes from your earlier chats and keeps on your machine under ~/.codex/memories/. They are off by default, separate from ChatGPT memory, and not shared with teammates (OpenAI docs, read 28 October 2026).
How do I turn on Codex memories?
Add memories = true under [features] in ~/.codex/config.toml, or turn on Enable Codex memories in the ChatGPT desktop app under Settings, Personalization. Use /memories in a chat to keep that one chat out, without changing the global setting.
Where are Codex memories stored?
In the memories folder under your Codex home, ~/.codex/memories/ by default: MEMORY.md, memory_summary.md, raw_memories.md, rollout_summaries/ and skills/. From Codex CLI 0.165.0 (27 October 2026), a profile that turns memories on for the first time uses the v2 layout in ~/.codex/memories_v2/.
Does Codex share memory with ChatGPT?
No. OpenAI says ChatGPT web uses ChatGPT memory, while local Codex clients use a separate local memory store. The exception is Computer History, a macOS feature of the ChatGPT desktop app that writes memories both ChatGPT and Codex can use.
Can Codex read Claude Code memory?
Only through /import. In a local Codex CLI session, /import copies a Claude Code setup across, mapping its project memories to Codex memories and its instruction files to AGENTS.md. In the CLI it is a one-off copy; the desktop app can keep imports in sync.