When a ChatGPT project lasts weeks or months, the problem is not always that the model simply “forgets.”

Something more dangerous can happen: a new conversation reconstructs an almost-correct version of the project and continues reasoning perfectly well from incomplete history.

We ran into this ourselves. We were about to change part of a project when we recovered an older continuity record and discovered that:

• a decision we thought was new had already been made,

• a capability we thought was missing already existed,

• and something we were about to modify was supposed to remain frozen.

The reasoning wasn’t bad. The available context was incomplete.

So we built a simple process called Continuity that extracts the current operational state of a conversation and carries it into a new one: active objectives, decisions still in force, constraints, completed work, unresolved items, dependencies and critical references.

It doesn’t try to make the AI remember everything. It preserves what is still necessary to continue working correctly.

After repeated testing on long-running projects, the effect has been surprisingly large for such a simple mechanism.

This seems like a useful native ChatGPT capability: not unlimited memory, but explicit operational continuity between conversations.

There is a free working implementation here for anyone who wants to test the idea:

I’d be interested to know whether others working on long-running ChatGPT projects have encountered the same failure mode.