Codex memory still on GPT-5.6 Luna? Pin newer extract and consolidation models

Local Codex memories do not have to stay on the background models Codex picks by default. You can override the per-chat extraction model and the global consolidation model in ~/.codex/config.toml.

#What changed

Community reports, including @daniel_mac8, say Codex memory defaults to GPT-5.6 Luna for extraction and GPT-5.6 Terra for consolidation. That default pair is a field report. What OpenAI's docs do confirm is the override: memories.extract_model and memories.consolidation_model.

To try newer models with memories on, set both keys under the right tables in config.toml (bare keys alone are incomplete):

[features]
memories = true

[memories]
extract_model = "gpt-6-luna"
consolidation_model = "gpt-6.1-sol"

Use model IDs your account can actually call; if a slug is unavailable, swap in one you can use. Restart or start a new Codex session after saving.

#Who is affected

Anyone using local Codex memories in the CLI, IDE extension, or desktop app. The feature is off by default until enabled in Settings → Personalization or via [features] memories = true. ChatGPT web memory is separate and is not controlled by these keys.

#What to do now

1. Enable Codex memories if you want them.
2. Set extract_model and consolidation_model under [memories] to models you prefer and can access.
3. Keep durable team rules in AGENTS.md; treat memories as recall, not the only source of standing instructions.
4. Memory generation can skip when your remaining rate-limit percentage is low, so stronger memory models are not a "free" usage hack.

#What is not confirmed

• That GPT-5.6 Luna/Terra remain the live defaults for every build (field report, not spelled out on the Memories help page).
• Whether pinning gpt-6-luna / gpt-6.1-sol always improves memory quality for every workflow.
• The exact usage cost of background memory passes on stronger models.

#Sources

• Field report: @daniel_mac8 on X
• Official docs: Codex memories (extract_model / consolidation_model)
• Config reference