AWS puts a clock on agent memory: 90 days for episodic, no expiry for procedural
AWS has published guidance on lifecycle policies for the memory of long-running agents. Memory is split into three kinds with a separate retention period for each, and a dedicated handler deletes every record belonging to one user for right-to-be-forgotten requests. It also gives a cost figure: 100,
AWS has published guidance on how long a long-running AI agent should keep what it remembers. The post went up on 4 September; it splits agent memory into three kinds, proposes a separate retention period for each, and ties the whole thing to working code. It looks technical, but it is a retention policy question, which makes it a legal and cost question.
Three kinds of memory, three periods
- Episodic, the conversation records: expire these first, default 90 days.
- Semantic, extracted facts and preferences: 6 to 12 months.
- Procedural, learned workflows and tool-use patterns: the longest; AWS suggests considering no expiry at all.
Three policies
The first is expiry by time to live, defaulting to 90 days for episodic records and 30 to 60 days for summaries. The second is relevance decay: creation recency is weighted 0.4, last-access recency 0.35 and access frequency 0.25, with the access count saturating at 50. The third is model-assisted consolidation, merging related memories into a single compact entry.
The pruning window in the post's table varies by the agent's job: 7 days for a real-time support bot, 21 for sales and onboarding, 45 for a general assistant, 90 for an IT helpdesk, and 180 for a legal and compliance advisor.
Compliance and cost
The post defines a dedicated handler that deletes every memory belonging to one user for right-to-be-forgotten requests, and leaves the audit trail in CloudTrail. On cost it gives one figure: 100,000 memories can reach 50 to 100 dollars a month. The reference build is Step Functions, five Lambda functions and an EventBridge rule firing nightly at 02:00 UTC; the code sits in the aws-samples repository.
For the corporate reader
Agent memory holding personal data is a record system under data protection law. Putting an agent into production without a retention period, a deletion procedure and an audit trail means accumulating an archive that never shows up in your data inventory. The AWS table is a good starting point for that conversation; the periods still have to be mapped onto your own retention policy.
Sources
- Primary source: AWS Machine Learning blog, 4 September: aws.amazon.com/blogs/machine-learning
- Sample code, aws-samples repository
