reflexion

# reflexion Blind retry repeats the same mistake. After a failure, turn the failure signal into a written lesson, then retry WITH that lesson in front of you. Verbal self-reflection as memory between attempts. ## Why this exists (evidence) - Reflexion (Shinn et al., arXiv:2303.11366): agents that verbally reflect on failure feedback and keep it as episodic memory across attempts get large gains, e.g. up to 91% pass@1 on HumanEval vs 80% for the prior best at the time, and big jumps on decision-making/reasoning tasks, without finetuning. - Mechanism: the failure (stack trace, failed test, wrong result) is signal; a concrete "what went wrong + what to change" note steers the next attempt instead of resampling the same path. ## When to use - The 2nd (and later) attempt at any task that just failed: failing test, runtime error, wrong answer, rejected approach. - A loop that is about to retry: insert a reflection step first. - NOT on the first attempt (nothing to reflect on yet) and not when the fix is obvious (just fix it). ## The loop 1. **Capture the failure precisely:** exact error / failing assertion / wrong output. Quote it, do not paraphrase. 2. **Reflect (short, concrete):** WHY did it fail? What specific assumption was wrong? What is the ONE change that addresses the root cause (not the symptom)? 3. **Record the lesson** as a one-line note kept in context (e.g. "the API returns cents not dollars -> divide by 100"). 4. **Retry** with the lesson explicit. Change the approach the lesson implies, do not just rerun. 5. **If it fails again:** add a new lesson; after 2-3 failed reflections on the same root, STOP and escalate (you are stuck in a local minimum) rather than looping. ## Persist across sessions For recurring failures, write the lesson to memory (see recall / memory-persistence) so a future session does not relearn it. That is the "episodic memory" part of Reflexion. ## Composes with - `systematic-debugging`: reflexion is the between-attempts memory; debugging is the within-attempt method. - `testsmith` / `adversarial-verify`: produce the failure signal reflexion learns from. - `self-consistency`: when reflection is uncertain, sample a few candidate root causes and take the consistent one. ## Honest limits - Garbage reflection = garbage retry. The lesson must name a ROOT cause, not "try harder". - Cap the loop (2-3 reflections); reflexion does not rescue a fundamentally wrong approach, it just stops you repeating one mistake. - Reported numbers are from the paper's benchmarks; measure your own.