Memory should be time-aware and shared consistently across Codex and ChatGPT

Summary

If Codex and ChatGPT are moving toward a more unified product experience, their memory systems should be consistent and shared across surfaces. A merged experience without shared memory will feel fragmented, and a shared memory system that treats old and recent memories as equally valid can produce incorrect or unsafe behavior.

Problem

ChatGPT memory currently appears to treat memories from very different points in time as if they have equal relevance. Old context can continue to strongly guide recent conversations even when it may no longer be true.

For example, if a user previously discussed an injury or health condition a long time ago, the assistant may keep assuming the user is still in that state in later conversations, even though significant time has passed. This is not rigorous: time elapsed should affect whether a memory is still considered current.

This becomes more important if Codex and ChatGPT are expected to converge or share more user context. Product-level memory should not just be shared; it should also be time-aware, scoped, and revalidated when stale or sensitive.

Expected behavior

• Codex and ChatGPT should have a coherent memory model if their experiences are being merged.
• Memories should include temporal metadata and should not be treated as equally reliable forever.
• Old memories should decay in relevance unless refreshed by recent interactions.
• In time-sensitive or safety-sensitive domains, such as health, the assistant should avoid assuming old information is still true.
• The assistant should explicitly re-check stale memories with the user when they materially affect an answer.
• Users should be able to inspect, update, expire, or archive memories by age and scope.

Suggested requirements

1. Add recency weighting or expiration semantics to memory retrieval.
2. Show timestamps or last-confirmed dates in user-visible memory management.
3. Distinguish persistent preferences from temporary life-state facts.
4. Require confirmation before applying stale memories in sensitive contexts.
5. If Codex and ChatGPT share product identity or workflows, make the memory layer interoperable across both surfaces.

Why this matters

A unified Codex + ChatGPT experience depends on coherent personal context. But inaccurate stale context can be worse than no memory at all, especially when it affects health, work assumptions, project state, or long-running user preferences.