Long Memory — Storage
Configure record expiration and conflict resolution for long memory.
The Storage tab controls how long memories persist and how the system handles semantically similar or contradictory information.
Time to Live (TTL)
Set an automatic expiration for memory records. Records older than the specified number of days are permanently deleted from both Qdrant and PostgreSQL.
How it affects your pipeline:
- Storage cost — Shorter TTL reduces vector database storage and associated costs
- Search relevance — Old memories may contain outdated information; TTL ensures automatic cleanup
- Compliance — Meet data retention requirements by enforcing maximum storage duration
- Leave empty — Records persist indefinitely (subject to manual cleanup only)
Recommendations by use case:
| Use Case | Recommended TTL |
|---|---|
| User preferences | No expiration — preferences are long-lived |
| Customer support history | 90 — 180 days — balance context with storage |
| Temporary project knowledge | 30 — 90 days — align with project lifecycle |
| Regulatory compliance | Match your retention policy (e.g., 7 years) |
Irreversible: Once records expire, they cannot be recovered. Test with a short TTL initially to verify your pipeline handles expiration gracefully.
Conflict Resolution Strategy
When a new memory is semantically similar to an existing one, the conflict resolution strategy determines how to handle the overlap. This prevents duplicate or contradictory memories from polluting search results.
Newer Wins
Replaces the existing memory with the newer information.
Best for: User preferences, settings, and facts that change over time. If a user previously said they prefer Plan A but now says Plan B, the latest statement is the correct one.
Effect on pipeline: Keeps memory count stable but may lose historical context about previous preferences.
Keep Both
Stores both the old and new memories as separate records.
Best for: Factual knowledge, research notes, and domains where multiple perspectives are valid. Both versions remain searchable, and the agent can consider historical context.
Effect on pipeline: Memory count grows over time. Search results may include contradictory information — your agent prompt should handle this.
Confidence Based
Compares the confidence scores of both memories and retains the one with higher confidence.
Best for: Knowledge bases where information quality varies. High-confidence sources (verified facts) override low-confidence ones (casual mentions).
Effect on pipeline: Maintains information quality but requires the summarization pipeline to assign meaningful confidence scores.
Tip: Start with "Keep Both" during development to observe how conflicts arise in your domain, then switch to a more aggressive strategy once you understand the patterns.
Next Steps
Proceed to Summarization to configure automatic memory consolidation.