How Long Should AI Remember a Customer

Retention by Memory Type

Episodic memories (specific interaction records) are most valuable in the days immediately following the interaction and lose relevance rapidly. The details of a chat conversation from last Tuesday are highly relevant this week, moderately relevant next month, and largely irrelevant six months from now. Set retention at 30 days for detailed records and 180 days for consolidated summaries. The consolidation process should extract durable facts from episodic memories before they expire, promoting important information to semantic storage.

Semantic memories (factual knowledge about the customer) remain relevant much longer because they represent stable characteristics. A customer's programming language, industry, company size, and primary use case change infrequently. Set retention at 12 to 24 months, with automatic renewal when the customer interacts and the information is confirmed. If a customer does not interact for 24 months, their semantic memories should expire because the information may no longer be accurate, and the customer may no longer be active.

Preference memories (how the customer likes to interact) have the longest useful lifetime because communication preferences are among the most stable customer attributes. A customer who prefers concise, technical explanations today is likely to prefer them next year. Set preferences to persist indefinitely but with a slow confidence decay of 10 to 20% per quarter. This means a preference learned from strong evidence a year ago is still active but with reduced confidence, so recent contradictory signals can override it more easily.

Procedural memories (what worked and what did not) have variable retention depending on the product's change rate. For fast-moving software products, troubleshooting procedures from a year ago may reference deprecated features or outdated configurations. Set retention at 90 to 365 days depending on how frequently your product changes. For stable products with infrequent updates, longer retention is appropriate.

Factors That Should Adjust Retention

Customer relationship value can justify longer retention for high-value accounts. An enterprise customer with a six-figure annual contract warrants a more complete memory profile than a free-tier user, because the cost of that customer having to re-explain themselves is higher. This does not mean storing more data, but keeping the existing data available for longer.

Regulatory requirements may mandate specific retention periods. GDPR requires that personal data not be kept longer than necessary for its stated purpose, which sets an upper bound. Industry-specific regulations (financial services, healthcare) may set minimum retention periods for records of customer interactions. Your retention policies must satisfy both the maximum (privacy regulations) and minimum (industry regulations) bounds.

Customer activity level should influence retention. Active customers who interact monthly have their memories naturally refreshed through consolidation. Inactive customers who have not contacted support in 12 months may have stale memories that represent their situation a year ago, not now. Consider reducing retention for inactive customers or flagging their memories as "last confirmed X months ago" so the AI knows to verify before relying on them.

The "Never Forget" Anti-Pattern

Some implementations default to keeping everything forever. This seems safe ("we might need it") but creates three problems. First, storage and retrieval costs grow without bound, and retrieval quality degrades as old, irrelevant memories compete with current ones for the top retrieval slots. Second, privacy exposure grows because every stored memory is a piece of customer data that regulations apply to and that could be included in a data breach. Third, memory accuracy degrades because old information that was once true may no longer be, and the system has no way to distinguish current from outdated facts without active maintenance.

The right approach is to expire memories on a schedule, promote important information through consolidation before expiration, and trust that if the information is still relevant, it will be re-observed in future interactions and stored fresh. This keeps the memory store lean, accurate, and compliant.

Industry-Specific Retention Guidelines

SaaS and technology products change frequently, which means customer memories about product usage, feature configurations, and troubleshooting procedures become outdated faster than in other industries. Set episodic retention at 30 to 90 days and semantic retention at 6 to 12 months. Product version changes should trigger a review of technical memories, flagging any that reference deprecated features or outdated APIs. A customer's memory profile that says they "use the v2 API" is actively harmful if v2 was deprecated six months ago and they have since migrated to v3.

E-commerce has seasonal patterns that affect retention strategy. Purchase preferences from last holiday season are relevant when this holiday season approaches, suggesting a 12 to 15 month retention for preference and purchase context memories. Return and exchange patterns are relevant for 90 to 180 days (covering the typical return window and any extended warranty periods). Size and fit information persists as long as the customer relationship is active, because clothing sizes and brand preferences are durable knowledge that saves time on every future interaction.

Financial services often face regulatory minimum retention periods that conflict with privacy minimization principles. Financial interaction records may need to be retained for 5 to 7 years under securities regulations, while GDPR requires deleting personal data when no longer necessary. The resolution is to separate the regulated records (stored in compliance-controlled systems with mandatory retention) from the AI memory (stored with standard retention and deletable on request). The AI memory provides the conversational context that improves service quality, while the regulated records satisfy compliance requirements independently.

Healthcare has similarly complex retention requirements. HIPAA does not specify retention periods for all records, but state laws may require retention of medical records for 6 to 10 years. Again, the resolution is separation: clinical records stay in the EHR system with mandatory retention, while AI memory stores only non-PHI context (communication preferences, scheduling patterns, common question types) with standard retention policies that can accommodate deletion requests.

Implementing Retention in Practice

Set retention at the time of storage, not retroactively. Every memory should carry an expires_at field in its metadata, calculated from the retention policy for its category. A daily expiration job scans for memories past their expiration date and processes them through the same complete deletion pipeline used for customer erasure requests: remove content, embedding, graph connections, and cached references.

Before deleting an expired memory, run a consolidation check: does this memory contain information that should be promoted to a longer-lived category? An episodic memory expiring at 90 days might contain the observation that the customer switched from JavaScript to TypeScript, which is a semantic fact worth persisting for another year. The consolidation check prevents valuable knowledge from being lost just because its container has expired, while still enforcing the retention limit on the original record.

Provide retention override capabilities for specific business needs, but require justification and approval. If a support manager wants to extend retention for a critical enterprise customer going through a complex migration, they should be able to set a custom retention period with a documented reason. These overrides should be reviewed quarterly to ensure they are still justified, and they should never override a customer's explicit deletion request.