Heedback’s AI doesn’t just answer questions — it learns. Every knowledge base article you write, every correction an agent makes in a conversation, feeds back into the AI’s memory. Over time, responses become more accurate, more aligned with your tone, and more reflective of the nuances only your team understands.

This guide walks you through how AI Memory works, how to manage it, and how to deliberately train the AI so it becomes a reliable extension of your support team.

Prerequisites

Before working with AI Memory, make sure:

You have an active Heedback organization with a Pro plan.
Your AI provider is configured under Settings > AI Configuration (OpenAI or Anthropic).
You have at least a few knowledge base articles published — these form the foundation of the AI’s understanding.
Your account has admin or owner permissions.

How AI Memory Works

Heedback’s AI builds its memory from two primary sources:

Knowledge base articles: When you publish or update an article, the AI indexes its content. This is the structured foundation — product explanations, policies, troubleshooting steps, and FAQs.
Human corrections in conversations: When an agent edits, overrides, or replaces an AI-generated response in the inbox, Heedback captures that correction as a memory entry. The AI treats corrections as high-priority context for future, similar questions.

This dual-source approach means the AI starts useful from day one (via your articles) and gets progressively smarter through real-world usage.

Viewing Your AI Memory Entries

To browse what the AI has learned:

Navigate to AI > Memory in the left sidebar.
You’ll see a list of memory entries, each showing the original question or context, the correction or learned fact, and the date it was captured.
Use the search bar to filter entries by keyword — helpful when you want to check if the AI already knows about a specific topic.

Memory entries are organized chronologically, with the most recent corrections at the top. Each entry shows its source: either “Knowledge Base” or “Agent Correction.”

Editing and Deleting Memory Entries

Not every correction ages well. Product features change, policies update, and sometimes an agent’s correction was situational rather than universal. You can refine the AI’s memory manually:

Edit an entry: Click on any memory entry to open it. Adjust the stored content to make it more precise or general. For example, if an agent corrected a response for a specific customer’s edge case, you might generalize it so the AI applies the lesson broadly.
Delete an entry: If a memory entry is outdated or incorrect, delete it. The AI will stop using it as context immediately.

Think of memory management like pruning a garden — regular maintenance keeps the AI’s responses sharp and relevant.

Training the AI Through Corrections

The most powerful way to improve AI responses is through everyday conversations:

Let the AI draft a response in the inbox when a customer message arrives.
Review the draft. If it’s accurate, send it. If it’s off, edit the response before sending.
The correction is captured automatically. Heedback stores the original AI draft, your corrected version, and the customer’s question as context.
Future similar questions will incorporate the correction. The AI learns that your version is preferred for that type of question.

This feedback loop means your team doesn’t need to set aside time for “AI training sessions.” Training happens organically as agents do their normal work.

Tips and Best Practices

Be consistent with corrections. If two agents correct the same type of question differently, the AI receives conflicting signals. Align your team on standard responses for common scenarios.
Write knowledge base articles first. The AI performs best when it has structured articles to draw from. Corrections are powerful but work better as refinements on top of a solid article foundation.
Review memory monthly. Set a recurring reminder to scan memory entries. Remove anything tied to deprecated features, old pricing, or expired promotions.
Use corrections for tone, not just facts. If the AI sounds too formal or too casual, editing responses teaches it your preferred communication style — not just the right answer.
Don’t over-correct. Minor phrasing tweaks for every response create noise. Focus corrections on factual errors, missing context, or significantly off-tone responses.

Knowledge Base: The primary content source for AI Memory. Well-structured articles lead to better AI responses from the start.
AI Auto-Reply: Enable automatic AI responses in the widget and inbox. AI Memory directly improves auto-reply accuracy.
Support Inbox: Where agent corrections happen. Every edited AI draft in the inbox feeds back into memory.
Widget: Your customers interact with the AI through the embeddable widget — the front line where AI Memory makes a visible difference.