Long Memory — Compliance
Configure PII redaction, permanent facts, and heat scoring for long memory.
The Compliance tab controls data privacy, information permanence, and relevance scoring. These settings ensure your memory system meets regulatory requirements while keeping search results focused on what matters.
Privacy & Compliance
Auto-Redact PII
Automatically detects and redacts personally identifiable information before storing memories. This includes names, email addresses, phone numbers, credit card numbers, and other sensitive data patterns.
How it affects your pipeline:
- GDPR/HIPAA compliance — Sensitive data is masked before it reaches storage, reducing regulatory risk
- Search behavior — Redacted content appears as
[REDACTED]in stored memories and search results - Performance — PII detection adds processing time to each store operation
- Irreversible — Once redacted, the original PII is not recoverable from memory
Enable for production: If your agent handles user data, enable PII redaction. The small processing overhead is negligible compared to compliance risk.
Extract Permanent Facts
Identifies and marks certain facts as permanent — immune to heat decay and TTL expiration. These are critical pieces of information that should never lose relevance.
How it affects your pipeline:
- Critical preservation — User identity, account status, and core preferences stay accessible regardless of activity
- Search priority — Permanent facts maintain maximum heat score, appearing prominently in results
- LLM-driven detection — The system uses an LLM to identify which facts qualify as permanent based on content analysis
Typical permanent facts:
- User account type and subscription tier
- Core accessibility preferences
- Legal agreements and consent records
- Organizational role and permissions
Heat Scoring
Heat scoring assigns a relevance value to each memory that changes over time. This prevents stale information from polluting search results while keeping actively used memories prominent.
Enable Heat Decay
Activates the heat scoring system. When enabled, memories gradually lose relevance unless accessed.
How it affects your pipeline:
- Search quality — Cold memories score lower in search results, pushing relevant information to the top
- Natural forgetting — Mimics human memory by deprioritizing unused information
- No automatic deletion — Heat decay affects search ranking only; memories are not deleted (use TTL for that)
Decay Rate
Percentage of heat lost per decay period. Range: 1% to 50%.
How it affects your pipeline:
- Higher decay — Memories cool quickly; only recently accessed information stays prominent
- Lower decay — Memories retain relevance longer; broader historical context available
| Decay Rate | Memory Lifespan (at daily period) | Best For |
|---|---|---|
| 1% — 10% | Weeks to months | Knowledge accumulation, research |
| 10% — 25% | Days to weeks | Standard chatbots, support agents |
| 25% — 50% | Hours to days | Fast-moving contexts, news, ephemeral data |
Decay Period
How frequently heat decay is applied. Options: Hourly, Daily, or Weekly.
How it affects your pipeline:
- Hourly — Fine-grained decay for high-activity agents; heat changes are smooth and gradual
- Daily — Balanced approach; most common default
- Weekly — Coarse decay for low-activity agents; memories stay warm longer between checks
Effective decay calculation: A 10% daily decay means a memory loses ~63% of its heat after 30 days (compounded). Combine rate and period to tune your forgetting curve.
Access Boost
Percentage of heat added when a memory is accessed via search. Range: 5% to 100%.
How it affects your pipeline:
- Frequently accessed memories — Stay warm through repeated use, creating a positive feedback loop
- Discovery effect — When a memory surfaces in search, it becomes more likely to appear again
- Higher boost — Strong recency bias; recently searched topics dominate results
- Lower boost — Gentler reinforcement; broader memory distribution in results
Minimum Threshold
Memories below this heat score are excluded from search results. Range: 0% to 50%.
How it affects your pipeline:
- Noise reduction — Cold memories filtered out, improving result quality
- Higher threshold — Only very active memories appear; precise but may miss relevant older information
- Lower threshold — Broader search scope; includes older memories that might still be relevant
- Zero threshold — All memories searchable regardless of heat (effectively disables filtering)
Recommendations:
| Use Case | Recommended Threshold |
|---|---|
| Precision-focused agents | 20% — 30% |
| General-purpose assistants | 10% — 20% |
| Knowledge retrieval (broad) | 0% — 10% |
Tuning tip: Start with a 10% daily decay, 20% access boost, and 10% minimum threshold. Monitor search result quality and adjust based on whether your agent recalls too much or too little.
Next Steps
Proceed to API Setup to generate your API key and review available endpoints.