Agent Composer Malfunctions Over Time

[Novax]

After extended use of the Agent Composer in Cursor AI, the following issues occur:

1. Model Inaccuracy and Unintended Deletions:

• The model starts to misbehave — it stops accurately applying changes and begins deleting or modifying unrelated sections of code.
• This behavior appears after prolonged use within a single Composer session.

2. Composer Chat Not Saving Properly:

• Some Composer chats are either not saving at all or become corrupted.
• When returning to these chats, they are either:
  • Completely gone, or
  • Display only a few messages bunched together, missing the rest of the conversation history.

3. Loss of Transparency in AI Actions (Long Sessions):

• Initially, when the AI runs commands or edits, you can see its actions clearly within the Composer.
• After a while, these actions become opaque — the UI no longer reflects what the AI is doing, which leads to confusion and loss of trust in the editing process.

[Novax]

The ai model breaks and stops working in general in long composers

[lpalbou]

the active memory context is summarized at some point, losing some critical information