Top Observations

The Top Observations panel shows your highest and lowest-confidence learned patterns, along with quality flags. This helps you quickly identify your strongest knowledge assets and potential problem areas in memory.

Understanding Weight

Each observation has a weight score (0-1) that reflects:

1. Confidence - How certain the agent was when creating it
2. Recency - Recently confirmed observations rank higher
3. Feedback - Positive feedback increases weight; corrections decrease it
4. Corroboration - Observations confirmed by multiple agents rank higher

Formula (simplified):

weight = base_confidence × recency_decay × feedback_multiplier × log₂(1 + corroboration_score)

Top Observations (Highest Weight)

These are your most trusted learned patterns. Characteristics:

• High agent confidence (typically >0.8)
• Recently confirmed or actively used
• Positive feedback from users
• Often corroborated by multiple agents

Use for:

• Export for documentation or onboarding
• Reference in team discussions
• Cross-project knowledge transfer (validated patterns)

Example high-weight observation:

"Deploy requests should always run 'npm run build:verify' 
 before merge to prevent type errors in production."
weight: 0.92 | confidence: 0.95 | corroborated: 3 agents

Bottom Observations (Lowest Weight)

Low-weight observations may indicate:

1. Uncertain learning - Agent was tentative when creating it
2. Stale information - Pattern no longer relevant
3. Negative feedback - Marked as incorrect or misleading
4. Isolated - No corroboration from other agents

Flags to watch:

• Marked misleading - User explicitly flagged as wrong
• Marked outdated - Still may be informative for historical context
• Low corroboration - Not independently verified

Use for:

• Audit and cleanup (delete stale patterns)
• Identify knowledge gaps (why is this observation uncertain?)
• Training data analysis (which learning scenarios are unreliable?)

Quality Flags

The panel highlights observations with quality concerns:

🚩 Misleading

• Explicitly marked as incorrect by feedback
• Should be reviewed for deletion or archival
• May indicate a learning bug in that agent

⚠️ Outdated

• Still recorded but flagged as no longer current
• Useful for historical analysis
• May be archived per your retention policy

💡 Low Corroboration

• Generated by a single agent
• May be overly specific or contextual
• Consider validating manually before relying on it

📉 Downweighted

• Started with moderate confidence
• Received corrective feedback
• May have been superseded by a better observation

Common Patterns

Feedback Workflow

To improve weight accuracy:

1. Review the observation content
2. Assess whether it's currently accurate (Yes/No/Unsure)
3. Provide feedback:
   • ✅ Upvote - Confirm it's accurate (increases weight)
   • ❌ Downvote/Mark incorrect - It's wrong (decreases weight, flags as misleading)
   • 🔄 Mark outdated - No longer current (preserves for history but lowers weight)

The platform uses your feedback to retrain observation weights and improve future learning.

API Reference

Endpoint: GET /api/memory/analytics/top-observations

Query Parameters:

• projectId (optional) - Filter to a specific project
• since (optional) - ISO-8601 date lower bound (default: 30 days ago)
• limit (optional) - How many top/bottom to return (default: 20, max: 100)

Response:

{
  top: Array<{
    id: string
    agentId: string
    content: string          // truncated to 300 chars
    weight: number           // feedback weight from observations table
    band: "high" | "neutral" | "low"  // high >2.0, low <0.5
    isMisleading: boolean    // true if id appears in any session_diagnostics.misleading_observation_ids
    createdAt: string        // ISO-8601
    metadata: Record<string, unknown>
  }>
  bottom: Array<{ /* same shape as top */ }>
  misleadingIds: string[]   // union of all misleading IDs from session_diagnostics
  projectId: string | null
  since: string
}

Example:

curl -X GET "https://api.rensei.ai/api/memory/analytics/top-observations?limit=5" \
  -H "Authorization: Bearer rsk_..."

Response:

{
  "top": [
    {
      "id": "obs_abc123",
      "agentId": "agent-frontend",
      "content": "Use React hooks for state management",
      "weight": 2.34,
      "band": "high",
      "isMisleading": false,
      "createdAt": "2026-05-10T08:15:00Z",
      "metadata": { "type": "coding-pattern" }
    }
  ],
  "bottom": [
    {
      "id": "obs_xyz789",
      "agentId": "agent-ops",
      "content": "Deploy on Tuesdays",
      "weight": 0.12,
      "band": "low",
      "isMisleading": true,
      "createdAt": "2026-04-01T14:20:00Z",
      "metadata": {}
    }
  ],
  "misleadingIds": ["obs_xyz789"],
  "projectId": null,
  "since": "2026-05-01T00:00:00Z"
}

Best Practices

1. Review top observations periodically - Confirm they still align with current practices
2. Provide feedback proactively - Upvoting correct observations improves the feedback loop
3. Investigate misleading patterns - Understand why they were learned to prevent recurrence
4. Track weight trends - Use export data to see how weights evolve over time
5. Validate before sharing - High-weight doesn't guarantee safety-critical correctness

Related Pages

• Memory Health - Deduplication ratios and storage metrics
• Drift Alerts - Contradictions that may lower weight
• Coverage Heatmap - Where these observations come from
• Memory Export - Download high-weight observations for sharing

Rate Limits

The top-observations API enforces a 100 req/min quota per organization.