How Codex memories work
With memories off, a new chat does not pull in what Codex learned in earlier chats. You can still resume an earlier session, for example with codex resume, and AGENTS.md and the repo files are always there. With memories on, Codex turns eligible earlier chats into local memory files: generated summaries plus the evidence they rest on. In a later chat Codex can recall them. Memory is built in the background, not right after every chat.
| Question | Answer (Codex docs, November 2026) |
|---|---|
| On by default? | No, local Codex memories are off by default |
| Turn on | ChatGPT desktop app: Settings > Personalization > Enable Codex memories; or memories = true under [features] in config.toml |
| Where it lives | In your Codex home directory, ~/.codex by default (a new profile uses memories_v2/) |
| Per-chat control | /memories, in the desktop app, the Codex TUI and the Codex CLI |
| Clear | Delete Codex memories, for the selected machine |
| Secrets | Redacted |
| Team sharing | Not described |
Turn it on in three steps
- Open the ChatGPT desktop app, go to Settings > Personalization and turn on Enable Codex memories. Or add
memories = trueunder[features]in ~/.codex/config.toml. - Work a few sessions as usual. Codex fills memory from earlier chats.
- Look in ~/.codex to see what was stored, and use
/memoriesto limit it in a given chat.
The misconception: memories are where team rules go
People who turn memories on often hope Codex will pick up the team’s conventions by itself. OpenAI says the opposite: treat memories as a helpful recall layer, not as the only source for rules that must always apply, and keep required team guidance in AGENTS.md or checked-in documentation. A memory is a summary Codex wrote itself, on one laptop, from one person’s chats. Nobody reviews it automatically, it is not automatically shared with teammates, and it is not in git. How AGENTS.md does travel with the repo is in What is AGENTS.md and how does Codex read it?.
The docs mention one exception to local storage: ChatGPT Work uses account and workspace memory settings. How that relates to a team’s Codex work, the page does not say.
What this means for a team
Say a developer finds out with Codex that the integration tests only pass with a certain migration order. Codex remembers that, on their laptop. A colleague asking Codex the same question gets nothing. If that colleague uses Claude Code, same result: Claude Code has its own auto memory, and Anthropic calls it “machine-local”, not shared across machines or cloud environments. On a new laptop, the old machine’s local memories are not there automatically.
So local memories do not automatically become team knowledge. What does get shared is whatever someone commits to AGENTS.md, documentation or code. A workable handoff: inspect the memory in ~/.codex, put the rule that holds into AGENTS.md, and have a colleague confirm in a new session that Codex loads it. If you run Codex and Claude Code side by side, read Codex vs Claude Code. How to capture company knowledge for AI is in How to give AI agents your company context.
Where Bombos fits
This is exactly the gap Bombos closes. It is a plugin that gives Codex and the team’s other AI tools one shared memory: what one person learned and the rules the team agreed on are ready for the next, traceable, in the cloud or on your own server. One developer can try it first for their own agents, and bring the team along when it works.
Related questions
How do I turn on Codex memories?
In the ChatGPT desktop app under Settings > Personalization > Enable Codex memories, or in config.toml with memories = true under [features].
How do I clear what Codex remembers?
Choose Delete Codex memories and confirm. That resets the memory store of the selected machine; per chat you control memory with /memories.
Do secrets from my code end up in memory?
Codex redacts secrets from what it stores. Still review the memory files before sharing your Codex home directory or generated memories with anyone, as OpenAI also advises.
Does Claude Code remember what we did?
Claude Code has auto memory, on by default in local sessions, per repository under ~/.claude/projects. It is also machine-local: not shared across machines, colleagues or cloud environments.
