I frequently use the same ChatGPT conversation over multiple days rather than treating each chat as a single session.
Currently, adjacent messages can appear semantically continuous to the model even when they were actually sent many hours or days apart. This can cause ChatGPT to incorrectly assume that a user’s mood, situation, plans, progress, or references such as “today”, “now”, and “recently” are still unchanged.
For example, one message might be sent on August 11 at 10:21, while the next message is sent on August 12 at 12:30. They are adjacent in the conversation history, but more than 26 hours have passed in the real world.
The ChatGPT web client already appears to receive per-message creation timestamps. While debugging the frontend, I can observe message metadata such as:
create_time: 1786509027.507247
This corresponds to the actual creation time of the message.
My suggestion is not necessarily to inject a visible timestamp string before every message. Instead, the conversation/inference layer could preserve temporal metadata or derive information such as elapsed time since the previous message.
For example:
previous user message → 26 hours ago
This could help the model distinguish between:
- a continuation after 30 seconds
- returning to a topic the next day
- resuming a conversation several weeks later
This seems especially useful for long-running conversations involving software projects, work progress, personal plans, learning, journaling, or any situation where the user’s state can change over time.
The key idea is:
message adjacency should not automatically imply temporal continuity.
Since timestamp metadata already appears to exist in the ChatGPT client data, exposing some form of temporal context to the model could substantially improve long-term conversational continuity.