Reflection & Learning (ERL)
Gestura includes an ERL-inspired reflection system for low-quality turns. At a user level, this means the app can notice that a response was weak, generate a structured correction, and sometimes reuse that correction later.
What ERL means here
The runtime follows a simple loop: attempt the task, evaluate quality signals, reflect on what went wrong if the turn was poor, and use that reflection to improve the next answer. In Gestura, this is an optional feature rather than a hidden always-on behavior.
What reflection can do
- Generate a short explanation of what failed or felt weak.
- Suggest a corrective strategy for the next attempt.
- Guide a same-turn retry without replaying tool side effects.
- Promote strong reflections into longer-term memory for future retrieval.
Why the text-only retry matters
Reflection-guided retries are designed to improve wording or reasoning without blindly re-running shell commands, file writes, or other side-effectful actions. That makes the feature safer than naive automatic re-execution.
When to enable it
Reflection is most useful when you want better recovery on complex tasks and you accept a little extra latency or cost. If you want the leanest possible path, keep it off. If you want more self-correction on weak turns, enable it and tune the thresholds carefully.
What to tune
- Quality threshold: how weak a turn must be before reflection starts.
- Max injected reflections: how much prior corrective guidance can flow back into future prompts.
- Max retry attempts: how many same-turn revision attempts are allowed.
- Promotion confidence: how strong a reflection must be before it becomes durable memory.