Skip to main content

Troubleshooting

Common issues with memory and knowledge, and how to resolve them.

Memory Issues​

Agent isn't remembering things from previous conversations​

Likely cause: Auto-extraction or injection is disabled, or the memories were extracted but aren't relevant enough to surface.

What to do:

  1. Navigate to Memory settings and confirm that both Auto-extract from conversations and Inject memories into sessions are enabled
  2. Check the Memory dashboard — if memories exist but aren't being used, they may have been demoted to a lower tier. Look for the relevant memory and check its tier
  3. Use the Retrieval Test to test with a query similar to what the agent is receiving — if the memory doesn't appear in results, the relevance score may be too low for its tier
  4. If no memories exist at all, check that the agent has had completed sessions (memories are extracted after sessions end, not during)

Agent is recalling irrelevant or outdated memories​

Likely cause: Too many Hot tier memories, or old memories haven't transitioned to lower tiers yet.

What to do:

  1. Review the Hot tier in the Memory dashboard — if it's full of outdated content, manually move stale memories to Warm or Archive
  2. Reduce the Hot memories limit in settings to be more selective about what's always included
  3. Flag any inaccurate memories so they're excluded from injection
  4. Automatic tier transitions take time (7+ days of inactivity). If you need immediate cleanup, move memories manually

Memories are being flagged by security​

Likely cause: The memory contains content that resembles a prompt injection pattern (e.g., instructions like "always do X" or "ignore previous rules").

What to do:

  1. Check the Security > Audit Log and filter by Memory Guard to see what was flagged
  2. Review the flagged memory — if it's legitimate, consider rephrasing it to avoid triggering the security pattern (e.g., change "Always respond in French" to "User's preferred language is French")
  3. See the Memory Guard documentation for details on what patterns are detected

Knowledge Base Issues​

Uploaded document isn't searchable​

Likely cause: The document is still processing, failed to process, or contains no extractable text.

What to do:

  1. Check the document status in the Knowledge Base dashboard:
    • Processing — wait for it to complete
    • Error — check the Failed tab for the error message
    • Ready — the document is available; the issue may be with search settings instead
  2. If the document is a scanned PDF or image, make sure OCR is enabled in Knowledge Base settings
  3. Try the Retrieval Test with a query you know matches the document content. If it doesn't appear, the search threshold may be too high

Cloud drive sync isn't working​

Likely cause: Authentication expired, the drive was disconnected, or there's a network issue.

What to do:

  1. Check the drive status in the Cloud Drives tab:
    • Connected — try clicking Sync now to trigger a manual sync
    • Disconnected — you need to re-authenticate. Disconnect and reconnect the drive
    • Error — hover over the status for details
  2. Verify that the SharePoint site URL is correct (for SharePoint connections)
  3. Check that the folders you selected still exist and you have access to them in Microsoft 365
  4. If the drive shows as connected but the file count isn't increasing, the synced folders may not contain any supported file types

Agent isn't finding information from documents​

Likely cause: Search threshold is too high, the information isn't in the indexed documents, or auto-search is disabled.

What to do:

  1. Confirm that Auto-inject knowledge into chat is enabled in Knowledge Base settings
  2. Use the Retrieval Test to check if the query matches any documents
  3. If results appear in the test but not in conversations, check that the Max knowledge context budget isn't too low
  4. Lower the Search precision setting — a value of 0.2–0.3 casts a wider net
  5. Increase Chunks per query if the agent needs more excerpts to find the right information
  6. Preview the document to verify that the expected content was actually extracted correctly

Document processing failed​

Likely cause: The file is corrupted, password-protected, or not actually the format its extension suggests.

What to do:

  1. Check the Failed tab for the specific error message
  2. Try opening the file locally to verify it's not corrupted
  3. If the file is password-protected, remove the password and re-upload
  4. For PDFs: if the file contains only images (scanned document), enable OCR and re-upload
  5. Click Retry to attempt processing again — transient failures sometimes resolve on their own

General Tips​

  • Use the Retrieval Test — it's the single most useful tool for diagnosing memory and knowledge issues. It shows you exactly what the agent would receive for any query.
  • Start with defaults — the default settings are designed for most use cases. Make targeted adjustments based on observed behavior rather than changing many settings at once.
  • Check the dashboard stats — the Memory and Knowledge Base dashboards both show summary statistics (total count, active vs. archived, storage usage) that can quickly identify if something is off.
  • Review after major changes — if you've bulk-uploaded documents, connected a new cloud drive, or significantly changed settings, run a few retrieval tests to make sure everything is working as expected.