openclaw - 💡(How to fix) Fix Second correction of the same error type does not trigger behavioral change — accommodation gap

Summary

When the same type of error is corrected twice, the system records the correction but does not update its underlying task-handling approach. Repeated corrections accumulate in memory but the system does not improve its actual behavior.

Problem to solve

In daily production use, a consistent pattern is observed: when the same type of error is corrected a second time, it is treated as a new correction rather than a signal that the underlying approach needs to change.

For example: if task type X is handled incorrectly and corrected once — the correction is recorded. But when task type X appears again, the same incorrect approach is applied because the underlying task-handling logic was never updated. The correction counter reaches 2, triggering a "systemic failure" flag — but by then the user has already had to correct the same error twice.

This suggests the system has:

• Correction recording: new corrections are added to memory
• Missing accommodation: the underlying task-classification or approach pattern is not updated when a correction pattern repeats

Without accommodation, corrections accumulate but the system's actual behavior does not improve. The AI appears to learn (corrections recorded) but gets more verbose rather than more competent.

Proposed solution

Track correction patterns at the task-type level, not just the individual rule level:

1. When a correction matches an existing correction pattern (same task type), increment a pattern counter
2. When the same pattern reaches 2 occurrences, trigger a structured update to the relevant AGENTS.md section
3. This closes the loop: correction → pattern detection → accommodation → behavioral change

This does not require AI to autonomously decide what to change — it only requires tracking correction frequency at the task-class level and triggering a structured update when a pattern is confirmed.

Impact

• Affected users: anyone who corrects the AI more than once on the same type of task
• Severity: annoying — wasted time on preventable repeated errors
• Frequency: very common in daily use
• Consequence: the AI appears to learn but does not actually improve through experience