AWS published default retention math for agent memory: 90-day TTL, 45-day prune, and a three-weight decay score

AWS Machine Learning Blog·medium signalXBlueskyLinkedInCopy link

The AgentCore memory lifecycle post is the most concrete public guidance yet on when to delete an agent's memories. It sets TTL defaults by memory type (90 days episodic, 30-60 days for summaries, 6-12 months semantic, optionally indefinite procedural), a relevance decay score weighting creation recency 0.40, last-access recency 0.35 and access frequency 0.25 with a 0.3 threshold at 45 days, and LLM consolidation that folds five episodic memories into one semantic entry before deletion. Cost is roughly $0.01-$0.02 monthly at 1,000 memories and $50-$100 at 100,000, dominated by the Bedrock consolidation calls, and it recommends pruneDays by agent type: 7 for real-time support, 45 general, 180 for legal and compliance.