[ChatGPT Projects] Regression: chats no longer proactively use relevant Project Sources and over-rely on recent conversation context #48718

What subscription do you have?

ChatGPT Pro

Model / surface

GPT-5.6 Sol, ordinary ChatGPT Project chats (not Work).

Summary

A long-running ChatGPT Project has recently regressed in how ordinary Project chats use relevant Project Sources.

Until recently, Project Sources behaved much more like an active part of the working context. After a current Source was updated, a subsequent ordinary or even very short/abstract message could naturally lead the assistant to notice and use the changed Source without the user having to say “open this source” or identify where to look.

Since roughly mid-September 2026, the assistant increasingly over-relies on the most recent conversation messages instead. Relevant Project Sources are consulted much less proactively. Short references that previously resolved naturally against project context now often produce questions, guesses, or confident reconstruction from recent chat context.

This regression is not limited to cases where Project Instructions explicitly require a Source. It is observable more generally in normal Project conversation. Explicit instruction failures are simply the strongest reproducible form of the same problem.

The problematic behavior is:

relevant Project Source not consulted → recent conversation context treated as sufficient → model guesses/reconstructs state rather than retrieving canonical project context.

Previous behavior

The Project uses current Sources as canonical shared state. A typical workflow was:

1. Update a Project Source (for example, the current task list).
2. Continue the Project conversation normally.
3. The assistant would often proactively use the relevant updated Source and notice the changed state without being told exactly where to look.

This made Project Sources function as intended for an ongoing personalized workflow: the user did not have to act as a dispatcher telling the assistant which Source to open on every turn.

Current behavior

In recent weeks, the same kind of continuation increasingly behaves as if only the latest conversation messages exist:

• relevant Sources are rarely consulted unless explicitly prompted;
• short context-dependent messages are guessed from recent turns;
• the assistant may ask where to look even though the relevant Project Source already exists;
• an explicit “read/check the project context” may still select the wrong Source;
• after one interpretation is rejected, the model may generate additional hypotheses rather than reset and retrieve canonical Project Sources.

This is a material behavior change in a workflow that previously worked substantially more reliably.

Strong reproduction — September 27, 2026

This Project also has a planner workflow where Project Instructions explicitly require several inputs before a daily plan is proposed:

• calendar;
• weather;
• a chronology/context Project Source;
• the current general task-list Project Source;
• and, on workdays, the current work-task-list Project Source.

On September 27:

1. The planner checked calendar and weather.
2. It then started constructing the day without attempting to retrieve the mandatory general task-list Project Source.
3. When challenged about whether the list had actually been read, it did not immediately report that it had not.
4. Instead, it tried to reconstruct the missing state from chronology, old daily plans, thematic files, and nearby context.

This is an explicit-instruction example of the broader regression: the model behaves as though recent conversation context is sufficient and does not reliably initiate Project Source retrieval.

This is not simply a Project Memory expectation

The workflow was specifically redesigned to reduce dependence on cross-chat memory.

• Long-term context is periodically materialized into a dedicated Project Source.
• Current task lists are maintained as current Project Sources.
• Only the short interval since the latest context snapshot depends on recent planner-chat history.

The assistant is therefore not being asked to remember an entire long-running Project. It is being asked to make normal proactive use of a small set of canonical Project Sources, as it previously did more reliably.

Expected behavior

Project chats should treat relevant Project Sources as active project context, not as material that is effectively ignored until the user explicitly names a Source.

In particular:

1. Relevant Project Sources should be consulted proactively when needed to resolve the user's request or a short contextual reference.
2. Explicit workflow instructions requiring named Sources should reliably trigger retrieval before output is produced.
3. If the model lacks sufficient project context, it should retrieve before guessing.
4. After an interpretation is rejected, it should prefer retrieval/reset over a chain of new unsupported hypotheses.
5. The user should not need to repeatedly say “read”, “search”, or identify which Project Source contains context the Project already provides.

Actual behavior

The assistant increasingly treats recent conversation context as sufficient, silently omits relevant Project Source retrieval, and may confidently reconstruct project state instead.

Regression timing

The proactive use of Project Sources appears to have become noticeably less reliable around mid-September 2026, without an intentional change to this Project workflow or its instructions.

There are other recent first-party Project/source reports in this repository involving source availability and synchronization. Those may be related infrastructure problems, but they are separate from this issue: this report is about the model/orchestration layer failing to initiate or prioritize relevant Project Source retrieval even when the Source itself may be available.

A separate Project Source access failure observed during investigation is being tracked separately so it does not obscure this orchestration regression.

Privacy

The Project contains personal material, so private Source contents and screenshots are not included publicly. The reproduction above describes the control flow without exposing that content.