Skip to main content

Correction Memory

Correction Memory helps your AI agent avoid repeating the same mistake.

When enabled, InstaMCP can remember approved corrections and use them in future conversations. This lets your integration improve over time without retraining a specific model provider.

Why It Matters​

AI agents can sometimes answer with the wrong policy, tone, business detail, or workflow. Correction Memory gives your team a way to teach the AI agent what the correct behavior should be next time.

Use it for things like:

  • Correcting a wrong answer about your business, service, or catalog.
  • Teaching the AI agent the preferred company wording.
  • Preventing repeated mistakes in support flows.
  • Capturing business-specific rules that are not obvious from the original prompt.
  • Keeping learned corrections useful even if you switch models later.

Model-Agnostic Learning​

Correction Memory is not tied to one LLM provider or one model.

That means:

  • You can change the integration's model later.
  • Your accepted corrections can still guide future answers.
  • You do not lose all AI agent learning just because you move from one model to another.

Think of Correction Memory as integration-level learning that travels with your InstaMCP integration configuration.

When to Enable It​

Enable Correction Memory when:

  • The integration is used by real customers or internal teams.
  • You expect the AI agent to improve from feedback.
  • Your business has special rules, policies, or wording.
  • You plan to test or change models over time.

You may leave it off for early experiments, demos, or temporary integrations where you do not want the AI agent to remember corrections.

How to Enable Correction Memory​

  1. Open Integrations.
  2. Select the integration you want to improve.
  3. Open the integration configuration.
  4. Enable Correction Memory.
  5. Save the integration.
  6. Test the integration with a question where a correction should apply.

Depending on your workspace, you may also be able to choose whether Correction Memory uses the same model as the integration or a separate correction model.

How Teams Should Use It​

Use corrections as durable guidance, not as one-off chat replies.

Good correction:

When a customer asks about refunds, explain that refunds are available within 14 days for unused plans and ask for the order email before creating a support ticket.

Weak correction:

Say it better.

What Not to Store​

Do not use corrections to store:

  • Passwords.
  • Private API keys.
  • One customer's private information.
  • Temporary one-off instructions.
  • Sensitive legal or compliance guidance without review.

Verification Checklist​

  • Correction Memory is enabled on the integration.
  • Your team knows when to add or approve corrections.
  • Test questions reflect the corrected behavior.
  • Corrections do not include private credentials or customer secrets.
  • Model changes are tested before production rollout.