What does memory mean in an AI application?
The Harvard Chan AI Glossary
In our classes, we often pause to explain the ideas behind AI and the language used to describe it. These posts grew out of those conversations. They’re for anyone who would like to understand AI better, with no technical background assumed.
Memory is information that an AI application stores and brings into later interactions to support continuity. It may include a user’s preferences, agreed project decisions or facts about an organization. In this sense, memory is usually a capability of the surrounding application, not evidence that the underlying model has permanently learned something new.
Imagine telling a writing assistant that a particular document should use British spelling. The application may save that preference and supply it when you return. The model—the trained component generating responses—can then apply it because the information is present in its current context, meaning the material available for this task.
Several kinds of record can be involved. Conversation history records earlier exchanges. Task state records progress on an unfinished job. A reference store contains documents or approved procedures. A preference record describes how a user wants work presented. Keeping these purposes distinct helps prevent an old conversation from being treated as current policy.
Storage alone is not useful memory. The application must find the relevant record, bring it into the current task and use it correctly. It must also handle corrections. If a project’s delivery date changes, retrieving both dates without identifying which one supersedes the other can create confusion rather than continuity.
Remembered information needs an appropriate scope. A preference for one project should not automatically govern another. Information about one person should not become available to someone else merely because the same application serves both. Retention—the length of time information is kept—and deletion are part of memory design, not administrative details added afterward.
Memory also differs from training, which adjusts a model’s internal numerical settings using examples. A chatbot appearing to remember your last conversation does not show that its model was retrained. Equally, a model’s broad knowledge from training is not an exact archive of your project.
A practical care-coordination example is a documented patient preference to be called after 5:00 pm, together with an open commitment to follow up on an abnormal lab result. Useful application memory brings those facts to the next authorized care coordinator, with their source and date, rather than forcing the patient to repeat them. It also allows correction and distinguishes a current preference from an outdated one.
If you would like to find out more about AI and improve your leadership skills, consider the AI in Health Care Certificate of Specialization from the Advanced Learning Academy at Harvard T.H. Chan School of Public Health. You can take an individual course or work toward the full certificate.
AI is a fast-moving field. This explanation reflects our class discussions and related material at the time of writing. Terminology, capabilities and tools may change.
For suggested additions or corrections, please email me at tpanch@hsph.harvard.edu.
Dr. Trishan Panch
Program Director, AI in Health Care
Harvard T.H. Chan School of Public Health — Advanced Learning Academy