Known Limitations

ACE is not a universal solution. The following conditions weaken its effectiveness or make it unsuitable.

​

Depends on Strong Reflector

ACE assumes the Reflector can distill meaningful experiences from traces.

• Weak Reflector → Playbook becomes noise or even misleading
• For some domains, any model struggles to distill useful insights → results are naturally mediocre

This shares the same root cause as Dynamic Cheatsheet's limitation: memory quality depends on the model itself.

​

Depends on Reliable Feedback

Feedback sources include:

​

Not All Tasks Need Long Playbooks

Playbooks emphasize detailed domain specifics, but for some tasks this is redundant.

​

Relationship to Unlearning

ACE's context is human-readable, naturally supporting:

• Selective forgetting for privacy/legal reasons
• Deletion of outdated/incorrect information identified by domain experts

This is also a unique advantage over weight adaptation (fine-tuning).

​

When to Choose Fine-Tuning Over ACE

• Long-term stable rules where training cost can be amortized
• Context cost exceeds weight training cost
• Requires implicit knowledge injection across tasks/prompts

ACE's core value is cheap, interpretable, instantly updatable—not replacing all weight updates.