We spent years teaching systems when to forget.
Delete old records. Minimize personal data. Define retention periods. Honor requests for erasure.
Now we are building AI agents specifically to remember.
That sounds useful… and it is.
An agent that remembers your preferences, previous decisions, customer context, ongoing tasks, or lessons from earlier interactions can provide a much better service than one that always starts from zero.
But persistent memory creates a new question:
What should an AI agent be allowed to remember and when should it forget?
AI does not remember everything. Different systems use memory in very different ways. But increasingly, agentic systems can maintain persistent information across sessions: user context, conversation history, goals, preferences, and more, which helps with consistency .
The Open Worldwide Application Security Project (OWASP) now treats agent memory as a distinct security surface and specifically warns about memory poisoning. This means malicious or incorrect information being stored in persistent memory and influencing future behavior. Its guidance recommends memory isolation, expiration limits, and reviewing sensitive information before it is stored.
So, memory is not simply a feature, it’s something that needs a lifecycle.
What exactly did the agent remember?
Imagine an AI agent supporting a customer-service employee.
During a conversation, it learns that a particular customer is considered high risk. Later, that information influences how the agent handles another request.
But a number of questions immediately come up:
- Where did that information come from?
- Was it verified?
- Is it still accurate?
- Was it supposed to become part of persistent memory in the first place?
- Who can correct it?
This is where memory becomes very different from a traditional database field. Employees may not even know which pieces of information an agent has retained, or which are influencing its future decisions.
A mistake can therefore survive the interaction in which it was made.
Worse, malicious information may survive too. OWASP describes attacks in which manipulated content persisted in an agent’s memory, which then affects later sessions.
Deleting the source may not be the end of it either.
Now, consider another scenario.
A customer asks for their personal information to be deleted. So, you remove it from the CRM.
But was some of that information copied into an agent’s persistent memory, context store, or another component used to personalize future interactions?
If so, deleting the original record may not automatically remove every copy created elsewhere in the AI architecture.
That matters because data-protection rules have not disappeared simply because AI has arrived. GDPR includes principles such as data minimization and storage limitation, as well as a right to erasure in defined circumstances. The UK’s Information Commissioner’s Office similarly stresses that AI systems should not keep personal data for longer than necessary simply because it could become useful later.
It’s not enough to ask if an agent can remember all of this, the question also includes whether or not it should remember something, for what purpose, and for how long.
Memory needs an owner
Before enabling persistent memory for an AI agent, organizations should be able to answer:
- What information can the agent retain?
- Which information must never become persistent memory?
- How long is memory retained?
- Who can inspect, correct, or delete it?
- What happens when the original source data changes or is removed?
- Can we identify which memories influenced an important action or decision?
This is not simply a technical configuration question.
It belongs to privacy, security, records management, risk, compliance, and business ownership.
Because an agent that remembers can become more useful over time. However, it can also accumulate outdated assumptions, sensitive information, and mistakes.
For years, organizations have worried about whether AI knows enough. With persistent AI agents, we may soon need to worry about the opposite: Does your AI know when it is time to forget?












