Feedback Loop
Steer the next answer with the last one
Most AI products collect feedback as silent telemetry. A real feedback loop closes immediately — the user picks a reason ('too long', 'missed the point'), sees the system write a preference, and the next answer demonstrably uses it. Feedback becomes an active control, not a future training signal.
Framing
The problem
Feedback in AI products is mostly telemetry — users vote, then nothing visible changes.
The pattern
Make feedback a structured loop: reason chips, a visible preference write, and the next answer demonstrating the change.
Why chat breaks here
Chat hides whether feedback influenced anything; users either give up or invent rituals to be heard.
Risks
Over-eager preference writes can swing the assistant toward instability; users need to edit and unpin.
Avoid when
The session is short and feedback would not survive long enough to be useful.
Use when
Feedback should visibly shape future answers — not just feed silently into telemetry.
DOPE evaluation
- Directability
- Pin, edit, or unpin any preference at any time
- Observability
- Each feedback entry produces a visible memory write the user can edit
- Predictability
- The next answer uses the preference; before/after is comparable on the same prompt
- Explainability
- Each preference is labelled with the feedback that produced it
In the wild
- ChatGPT Memory Summary + Sources (OpenAI) — The book icon under a response opens Sources: which memory shaped the answer, an explanation of why it was used, and a ••• correction. The memory summary itself is editable inline — type a change into it or highlight a line to fix it. No reason chips, and the write is retrospective rather than answer-by-answer.
- Claude Memory Write-Back (Anthropic) — Memory saves as individual Topics listed in Settings → Memory — open one to read, edit or delete it, and the fix applies to every conversation from then on. You can also tell Claude in chat to remember, change or forget something. On by default for Free, Pro and Max; the old Settings → Capabilities path is gone.
- Cursor Bugbot Learned Rules (Cursor) — Downvotes on Bugbot comments, replies explaining what was wrong, and "@cursor remember [fact]" on a PR become written rules listed by name, content and scoped paths in the Bugbot repository-rules dashboard, editable or deletable there. Candidates only go active once signal accumulates, and auto-disable on negative signal — the write is visible, not immediate.
FAQ
When should I use the Feedback Loop pattern?
Feedback should visibly shape future answers — not just feed silently into telemetry.
When should I avoid the Feedback Loop pattern?
The session is short and feedback would not survive long enough to be useful.
What problem does Feedback Loop solve?
Feedback in AI products is mostly telemetry — users vote, then nothing visible changes.
Why is chat the wrong fit for this?
Chat hides whether feedback influenced anything; users either give up or invent rituals to be heard.
Related patterns
- Extends: AI Memory — Feedback is the input; memory is the artefact it writes. The loop closes when the user sees the write and can edit it.
- Alternative to: Constraint Rules Engine — Both shape future runs — feedback through accumulated preference, rules through declared constraint. Soft vs hard.
- Often paired with: Autonomy Dial — Tune autonomy with feedback as the agent learns what kind of decisions you defer or reverse.