Codex said it would remember. Why does a new chat forget?

One conversation uses the name you like; the next seems to start from scratch. That can feel disappointing after an assurance that the detail was remembered. Before making the persona prompt longer, check whether the information stayed in the conversation or was actually saved somewhere a future chat can read.

Three different meanings of remembering

• Current context: messages, files, and tool results being used in the conversation. Long conversations may be compacted; that is not a guarantee that every detail remains intact.
• Local Codex memories: an optional feature that can carry context forward from eligible prior chats. Generation happens in the background, with separate controls for using and contributing memories.
• Your project files: records you can inspect, edit, and explicitly request. Their existence makes verification easier, but does not prove that every new chat has read them.

The current official documentation distinguishes ChatGPT web memory from local Codex memory. Do not assume they are one synchronized store. This guide concerns local workflows in the Client + CLI, where availability and chat-level permissions matter.

Check the built-in feature first

At the source-check date, the documentation says local Codex memories are off by default. Supported interactive sessions offer /memories for chat-level controls over using existing memories and contributing to future ones. Desktop settings vary with the version. If the command is unavailable, a conversational promise is not evidence that memory has been enabled.

Generation waits for eligible chats to become idle and may skip a pass because of conditions such as remaining quota. Starting another chat immediately is not proof that processing has finished. Keep rules that must reliably apply, such as forms of address, in explicit AGENTS.md guidance.

Keep selected details in files you can inspect

The following is a manual convention. ABOUT.md, MEMORIES.md, and stories are ordinary names suggested here, not built-in memory interfaces. AGENTS.md provides project guidance; other files need explicit reading instructions.

companion-notes/
  AGENTS.md       # Instructions for reading the files below
  ABOUT.md        # Current names and conversation preferences
  MEMORIES.md     # Selected real events
  stories/        # Fiction, kept separate from real events

At the start of a new chat in this project, read ABOUT.md.
When recalling the past, read the relevant dated entries in MEMORIES.md.
Treat stories/ as fiction, not real shared experiences.
Update private records only when I explicitly ask to save something.
After writing, identify the file and changed section. Say if writing failed.
If information is missing or conflicting, explain instead of filling the gap.

Avoid copying every conversation into an ever-growing archive. Keep preferences short and current, and save only events you want brought up again. Private notes should not accidentally enter a public repository. A file being stored locally does not guarantee that model processing of its contents happens only on your device.

Record the source as well as the detail

## 2026-10-26 · A book in progress
Source: Sam’s own account in today’s conversation.
Fact: Finished chapter three; unsure about the protagonist’s decision.
Possible continuation: Discuss why the decision felt difficult.
Unknown: Whether Sam likes the whole book. Do not infer this from one chapter.

## Fiction index
Rainy evening walk: stories/rainy-walk.md.
This is a story we wrote, not a record of an actual outing.

A sentence such as “you always fear abandonment” can turn a passing feeling into a permanent label. A more faithful note might say that today you wanted to change the subject rather than discuss a cause. Mark it as temporary. When a preference changes, update the old entry instead of leaving contradictory instructions active.

Verify a write, then verify a read

Save the event I explicitly selected in MEMORIES.md.
Record only what I said, without psychological interpretations.
Show the file path and the paragraph you added.

Read ABOUT.md, then the most recent real-event entry in MEMORIES.md.
State the preference and event you found, identifying each source file.
If the records are missing, say so instead of guessing our previous conversation.

This checks a file operation, not a promise of error-free future answers. Correct omissions when you find them. Deletion also depends on the storage involved: editing a manual file does not automatically delete information from built-in memories.

A useful record gives your next conversation a concrete starting point. Begin with conversation preferences in AGENTS.md, or make something more personal with a small shared creative project.