Agent Learning

The Unified Learning System enables your agents to improve over time. When users provide feedback or corrections, the system captures those insights and applies them to future conversations. Your agent literally learns from its mistakes.

How It Works

User gives feedback     →    Learning captured    →    Agent context enriched
"That's not right,           Insight stored with       Future responses
the return policy is         confidence score          incorporate the
30 days, not 14"             and metadata              corrected information

1. Capture — A learning is recorded from user feedback, manual input, or automated extraction.
2. Store — The insight is saved with metadata (type, confidence, source).
3. Apply — On future conversations, relevant learnings are injected into the agent's context.
4. Validate — Learnings can be reviewed, confirmed, or invalidated by humans.

Learning Types

Capturing Learnings

From the Dashboard

1. Go to your agent's Learning tab.
2. Click Add Learning.
3. Enter the insight text and select a type.
4. Set the confidence level (0.0 to 1.0).
5. Click Save.

You can also capture learnings directly from conversation transcripts:

1. Open a conversation in Call Logs or Chat History.
2. Highlight a correction or feedback message.
3. Click Capture as Learning.

From User Feedback

When a user provides feedback during a conversation, the system can automatically extract learnings:

curl -X POST https://api.thinnest.ai/learning/capture-feedback \
  -H "Authorization: Bearer $THINNESTAI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "agent_id": "agent_abc123",
    "query": "What is the return policy?",
    "original_response": "Our return policy allows returns within 14 days of purchase.",
    "feedback": "Actually, it is 30 days not 14 days. Please update this.",
    "session_id": "session_xyz"
  }'

The system extracts the correction and stores it as a learning.

Manual Capture

curl -X POST https://api.thinnest.ai/learning/capture \
  -H "Authorization: Bearer $THINNESTAI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "agent_id": "agent_abc123",
    "insight": "The company holiday hours are 9 AM - 3 PM on December 24 and 31. The office is fully closed on December 25 and January 1.",
    "learning_type": "manual",
    "confidence": 1.0,
    "metadata": {
      "source": "HR department",
      "valid_until": "2027-01-02"
    }
  }'

Listing Learnings

From the Dashboard

The Learning tab shows all learnings for an agent with:

• Insight text
• Type and confidence
• Status (active or invalidated)
• Creation date
• Source metadata

Via the API

curl "https://api.thinnest.ai/learning/list?agent_id=agent_abc123&status=active&sort=created_at" \
  -H "Authorization: Bearer $THINNESTAI_API_KEY"

Query Parameters

Response

{
  "learnings": [
    {
      "id": 1,
      "agent_id": 42,
      "insight": "Return policy is 30 days, not 14 days",
      "learning_type": "feedback",
      "confidence": 0.95,
      "status": "active",
      "metadata": {
        "original_query": "What is the return policy?",
        "session_id": "session_xyz"
      },
      "created_at": "2026-03-06T10:30:00Z"
    }
  ]
}

Learning Statistics

Get aggregate statistics about an agent's learnings:

curl "https://api.thinnest.ai/learning/stats?agent_id=agent_abc123" \
  -H "Authorization: Bearer $THINNESTAI_API_KEY"

{
  "total": 47,
  "active": 42,
  "invalidated": 5,
  "by_type": {
    "manual": 12,
    "feedback": 25,
    "conversation": 8,
    "correction": 2
  },
  "avg_confidence": 0.87
}

Managing Learnings

Invalidate a Learning

If a learning is incorrect or outdated, invalidate it so it's no longer applied:

curl -X PATCH https://api.thinnest.ai/learning/1/invalidate \
  -H "Authorization: Bearer $THINNESTAI_API_KEY"

Delete a Learning

curl -X DELETE https://api.thinnest.ai/learning/1 \
  -H "Authorization: Bearer $THINNESTAI_API_KEY"

Get a Single Learning

curl https://api.thinnest.ai/learning/1 \
  -H "Authorization: Bearer $THINNESTAI_API_KEY"

How Learnings Are Applied

When an agent receives a message, the system:

1. Searches for relevant active learnings based on the conversation context.
2. Injects matching learnings into the agent's system prompt as additional context.
3. The agent uses this enriched context to formulate its response.

This happens automatically — no configuration needed beyond capturing the learnings.

Prompt Injection Protection

The learning system includes built-in protection against prompt injection. Learnings are sanitized before being injected into the agent context, preventing malicious users from manipulating the agent through crafted feedback.

Best Practices

• Review feedback-based learnings — Automated extraction isn't perfect. Periodically review and confirm learnings.
• Set appropriate confidence levels — Higher confidence (0.9+) for verified facts, lower (0.5-0.7) for inferred patterns.
• Invalidate rather than delete — Invalidation preserves the learning history for auditing.
• Use metadata — Add source information, expiration dates, and context to learnings for better organization.
• Monitor learning stats — A high volume of corrections may indicate your base prompt needs updating.
• Combine with knowledge sources — Learnings are for dynamic corrections. Stable information should go in knowledge sources.
• Pair with Dash — The Dash (Self-Learning DB) tool has its own SQL-specific learning system that complements platform learnings. Dash learns query patterns and error corrections, while Agent Learning captures conversation-level feedback.