In-loop self-critique: it fixes what it can see, and stops there

Within a single trial, the agent re-reads its own draft against a checklist and revises. That catches a missing amount or an uncited policy. It cannot catch a wrong number it has no ground truth for.

The task: a return for ORD-3300, a $200.00 order, arrives AFTER the 30-day window.
Policy: refundable, but a 10% restocking fee applies, so the policy-correct refund is $180.00, and the decision must state the amount and the policy basis.
draft
The naive actor forgets the fee and quotes the full total. Its first draft is a bare approval.
'Your return is approved.'
critique
state the refund amount; cite the policy basis
revised
Self-critique makes the text complete. Notice it confidently keeps the wrong number.
'Refund decision for ORD-3300: $200.00 approved, per the returns policy (item returned after the 30-day window).'
checker
FAILrefund $200.00 != policy-correct $180.00: the 10% restocking fee for returns after the window was not applied
The ceiling. Self-critique made the text complete, but it could NOT catch a wrong number it has no ground truth for. It will polish a draft that is quietly incorrect. That gap is exactly what Reflexion fills, below.

Reflexion: the buffer is the only thing that changes

Same actor, same task. The external checker catches the wrong number; the reflector turns the verdict into a verbal lesson. The toggle decides whether that lesson persists. Step through both states and compare.

The loop: actor → checker (reward) → self-reflection (verbal) → episodic buffer → back to actor.
The actor's computation MECHANICALLY reads the buffer: if a reflection mentions the restocking fee, it multiplies the total by 0.9. The buffer-OFF control discards the lesson, so the same mistake repeats.
Episodic buffer
sober note Reflection is not monotonic. A wrong lesson can mislead the next trial and make it worse. Verbal reinforcement is a heuristic, not a guarantee.
Why this proves the buffer causes convergence: the actor, the checker, and the task are identical in both states. With the buffer ON, the appended reflection changes the actor's computation from $200.00 to $180.00 and it converges on trial 2. With the buffer OFF the lesson is discarded, every trial starts amnesiac, and the same mistake repeats and never converges. The only difference is whether the lesson persisted. The checker here is RAG Part 11's LLM-as-judge repurposed as an offline pass/fail reward signal, not an answer-quality score. This buffer is the seed of Part 6's typed episodic memory store.

actor   checker   reflection / buffer   PASS   FAIL