AWS introduces memory lifecycle policies for AgentCore
AWS outlines a three‑policy framework to automatically expire, score, and consolidate memories of long‑running Bedrock agents.
Original source published: September 4, 2026
A recent AWS blog post explains how unmanaged AgentCore memory can lead to outdated context and compliance concerns, citing examples where agents referenced resolved billing disputes or obsolete runbooks. To address this, the authors present a deployable architecture that runs nightly via Step Functions and an AWS CDK stack.
The framework defines three complementary policies. Policy 1 applies configurable TTLs—defaulting to 90 days for episodic memories, 30‑60 days for summaries, 6‑12 months for semantic memories, and no TTL for procedural memories. Policy 2 assigns a relevance score using an exponential decay formula that combines creation recency, last‑access recency, and access frequency, with default parameters such as pruneDays = 45, threshold = 0.3, and weights of 0.4, 0.35, and 0.25. Policy 3 leverages an Amazon Bedrock LLM to consolidate low‑scoring episodic entries into a single semantic record.
The post also offers recommended pruneDays values for different agent archetypes (e.g., 7 days for real‑time support bots) and provides the full implementation in a public GitHub repository, giving AI developers a ready‑to‑use pattern for managing agent memory as a managed resource.