Every team that trains an AI agent ends up doing the same thing. They test it, spot a wrong answer, and think: no, it should have said this.
Until recently, that thought had nowhere to go. You could fix the instructions, add a knowledge-base article, and hope the next conversation went better. Feedback Memory gives it somewhere to go: “that answer was wrong, here’s the right one” becomes a lasting memory for that agent, used on the very next turn.
Correct a reply in ten seconds
It all happens in the Playground, where you already test your agents.
- Hover a reply. Thumbs up and thumbs down appear beside it, and the whole reply is outlined, including every bubble of a split SMS.
- Pick a type. Thumbs down is a Correction: what the agent should have said. Thumbs up is a Reinforcement: a reply worth repeating.
- Write the fix and submit. You see the customer message and the agent’s reply for context while you write.
That’s it. The memory is live on the next turn, with no publish step and no review queue. You can find every memory later in Memory Hub, on the Feedback tab, where you can filter by agent, search and delete.
Playground conversations reset with each session, but feedback doesn’t. Training is the point.
Why corrections don’t wait for a lookup
Myna agents already have memory: general workspace memory, knowledge bases and conversation history. The agent fetches those when it decides it needs them, by calling a search tool.
For corrections, we deliberately didn’t do that. The reason is simple and a little uncomfortable: the corrections that matter most are exactly the ones the agent wouldn’t think to look up. If the agent knew its answer about same-day visits was wrong, it wouldn’t have given it.
So feedback works the other way round. On every turn, before the agent plans anything, Myna checks the customer’s message against that agent’s approved feedback and brings in the closest matches automatically. The agent doesn’t have to remember to ask.
| Scope | How it reaches a turn | |
|---|---|---|
| General memory | Workspace | Looked up when the agent decides to |
| Knowledge bases | Assigned to the agent | Looked up when the agent decides to |
| Conversation memory | Per contact | Looked up when the agent decides to |
| Feedback memory | Per agent | Brought in automatically, every turn |
Feedback is scoped to one agent, not shared across the workspace and not tied to a persona, because “what this agent should say here” is specific to that agent’s job.
Never store the mistake where it can be found
When you correct a reply, three things are saved: what the customer said, what the agent should have said, and the wrong answer.
Only the first two are used for matching. The wrong answer is kept for reference, but it’s left out of the searchable memory entirely. If it were searchable, the agent could retrieve its own mistake and repeat it. So the pairing that gets matched is always this kind of message, this right answer.
Your text is also stored exactly as you wrote it. Nothing rewrites, summarizes or “improves” a correction on the way in.
A threshold that matched nothing
How close does a customer message need to be before a correction applies? Our first answer was “very close”, and the first version was strict enough that it matched nothing at all.
The reason only showed up in testing. A short customer message like “can someone come today?” doesn’t look much like a stored pair of customer message plus full correct answer. Measured similarity came out far lower than intuition said it would. We recalibrated on real examples, comparing correct matches, near misses in the same domain, and unrelated messages, until corrections applied when they should and stayed quiet when they shouldn’t.
The before-and-after made the point. With feedback switched off, the agent promised a same-day visit that an operator had explicitly flagged as wrong. With it on, it didn’t.
That calibration turned out to be useful beyond feedback. We later reused it for the agent’s regular memory search too.
Authoritative on wording, advisory on actions
A correction changes what the agent says. It never changes what the agent is allowed to do.
When feedback matches, the agent is told to follow it over improvising, and to reuse its substance in the persona’s own voice rather than pasting it in. But it sits below everything that keeps a reply true:
- It never overrides what a tool actually returned. If a correction says “we have Saturday availability” and the calendar says otherwise, the calendar wins.
- It can’t turn on a tool the workflow step doesn’t allow, or get around pricing, truth and safety rules. Those work the way we described in how tools work inside a Myna agent.
- It’s never quoted as a system message, and the agent never mentions that training exists.
That’s what makes it safe to let anyone on the team leave a correction. The worst a bad correction can do is change some wording, and you can delete it in Memory Hub.
Who wrote it decides how much it’s trusted
A correction written by a person in the Playground is approved the moment it’s saved. That’s deliberate: if you’re training your own agent, waiting for approval would just slow you down.
The system is also built for corrections that don’t come from a person. Anything machine-generated is held as pending and can’t reach a live conversation until someone reviews it.
Train before you go live
In the Playground, feedback is always on, even for agents that haven’t been published yet. That means you can train an agent fully before a single real customer talks to it.
On live channels, Feedback Memory is a setting on each agent, and it’s on by default. Turn it off and the agent stops checking feedback entirely.
And it fails open. If anything goes wrong fetching feedback on a turn, that turn simply runs without it. A training feature should never be the reason a customer doesn’t get a reply.
The short version
- Correct a reply in the Playground, and it applies on the next turn.
- Corrections are brought in automatically, because the agent won’t look up what it doesn’t know it got wrong.
- The wrong answer is never searchable, so it can’t come back.
- Feedback shapes the wording and never overrides tools, prices or safety.
It’s the fastest way we’ve found to turn “it should have said this” into an agent that actually says it. Open the Playground, find a reply you’d change, and try it.
