AWS launches AgentCore memory lifecycle management

The post addresses a real operational pain point: long-running agents that remember everything can harm themselves. AWS observed failures like a support agent treating a resolved billing dispute as active, and another offering superseded deployment advice—both due to stale memories. The proposed solution separates memories into episodic, semantic, and procedural types, each with distinct retention needs. Episodic memories expire first, semantic memories last longer and consolidate into facts, and procedural memories are pruned only under the highest bar. The three policies—TTL expiration, relevance decay scoring, and LLM-based consolidation—work together in a nightly workflow. TTL runs first to avoid wasting compute, then scoring flags low-value memories for consolidation or pruning. The relevance formula balances recency and frequency, with weights adjustable per workload. Consolidation gives low-scoring memories a final chance by merging related episodes into durable semantic facts. The architecture uses AWS CloudTrail to track access, since AgentCore lacks a last-accessed field. Lower-volume agents may start with TTL and GDPR compliance alone, suggesting the full framework is for memory-intensive deployments.